rescue research paper

• Research article
• Open access
• Published: 05 February 2021

Increasing adoption rates at animal shelters: a two-phase approach to predict length of stay and optimal shelter allocation

• Janae Bradley 1 &
• Suchithra Rajendran   ORCID: orcid.org/0000-0002-0817-6292 2 , 3  

BMC Veterinary Research volume  17 , Article number:  70 ( 2021 ) Cite this article

38k Accesses

18 Citations

74 Altmetric

Metrics details

Among the 6–8 million animals that enter the rescue shelters every year, nearly 3–4 million (i.e., 50% of the incoming animals) are euthanized, and 10–25% of them are put to death specifically because of shelter overcrowding each year. The overall goal of this study is to increase the adoption rates at animal shelters. This involves predicting the length of stay of each animal at shelters considering key features such as animal type (dog, cat, etc.), age, gender, breed, animal size, and shelter location.

Logistic regression, artificial neural network, gradient boosting, and the random forest algorithms were used to develop models to predict the length of stay. The performance of these models was determined using three performance metrics: precision, recall, and F1 score. The results demonstrated that the gradient boosting algorithm performed the best overall, with the highest precision, recall, and F1 score. Upon further observation of the results, it was found that age for dogs (puppy, super senior), multicolor, and large and small size were important predictor variables.

The findings from this study can be utilized to predict and minimize the animal length of stay in a shelter and euthanization. Future studies involve determining which shelter location will most likely lead to the adoption of that animal. The proposed two-phased tool can be used by rescue shelters to achieve the best compromise solution by making a tradeoff between the adoption speed and relocation cost.

As the problem of overpopulation of domestic animals continues to rise, animal shelters across the nation are faced with the challenge of finding solutions to increase the adoption rates. In the United States, about 6–8 million dogs and cats enter animal shelters every year, and 3–4 million of those animals are euthanized [ 1 ]. In other words, about 50% of the total canines and felines that enter animal shelters are put to death annually. Moreover, 10–25% of the total euthanized population in the United States is explicitly euthanized because of shelter overcrowding each year [ 2 ]. Though animal shelters provide incentives such as reduced adoption fees and sterilizing animals before adoption, only a quarter of total animals living in the shelter are adopted.

Animal adoption from shelters and rescues

There are various places to adopt an animal, and each potential owner must complete the adoption process and paperwork to take their new animal home [ 3 ]. Public and private animal shelters include animal control, city and county animal shelters, and police and health departments. Staff and volunteers run these facilities. Animals may also be adopted from a rescue organization, where pets are fostered in a home or a private boarding facility. These organizations are usually run by volunteers, and animals are viewed during local adoption events that are held at different locations, such as a pet store [ 3 ].

There could be several reasons for the euthanization of animals in a shelter, such as overcrowding, medical issues (ex. sick, disabled), or behavioral issues (ex. too aggressive). The causes for the overpopulation of animals include failure to spay or neuter animals leading to reckless breeding habits and abandonment or surrender of offspring, animal abandonment from owners who are no longer able to take care of or do not want the animal, and individuals still buying from pet stores [ 4 ]. With the finite room capacity for animals that are abandoned or surrendered, overpopulation becomes a key challenge [ 5 ]. Though medical and behavioral issues are harder to solve, the overpopulation of healthy adoptable animals in shelters is a problem that can be addressed through machine learning and predictive analytics.

Literature review

In this section, we describe the research conducted on animal shelters evaluating euthanasia and factors associated with animal adoption. The articles provide insights into factors that influence the length of stay and what characteristics influence adoption.

Studies have been conducted investigating the positive influence of pre-adoption neutering of animals on the probability of pet adoption [ 2 ]. The author investigated the impact of the cooperation of veterinary medical schools in increasing pet adoption by offering free sterilization. Results demonstrated that the collaboration between veterinary hospitals and local animal shelters decreased the euthanization of adoptable pets.

Hennessy et al. [ 6 ] conducted a study to determine the relationship between the behavior and cortisol levels of dogs in animal shelters and examined its effect in predicting behavioral issues after adoption. Shore et al. [ 7 ] analyzed the reasons for returning adopted animals by owners and obtained insights for these failed adoptions to attain more successful future approvals. The researchers found that prior failed adoption had led to longer-lasting future acceptances. They hypothesized that the failed adoptions might lead owners to discover their dog preferences by assessing their living situation and the type of animal that would meet that requirement.

Morris et al. [ 8 ] evaluated the trends in income and outcome data for shelters from 1989 to 2010 in a large U.S. metropolitan area. The results showed a decrease in euthanasia, adoption, and intake for dogs. For cats, a reduction in intake was observed until 1998, a decrease in euthanasia was observed until 2000, and the adoption of cats remained the same. Fantuzzi et al. [ 9 ] explored the factors that are significant for the adoption of cats in the animal shelter. The study investigated the effects of toy allocation, cage location, and cat characteristics (such as age, gender, color, and activity level). Results demonstrated that the more active cats that possessed toys and were viewed at eye level were more likely to impress the potential adopter and be adopted. Brown et al. [ 10 ] conducted a study evaluating the influence of age, breed, color, and coat pattern on the length of stay for cats in a no-kill shelter. The authors concluded that while color did not influence the length of stay for kittens, whereas gender, coat patterning, and breed were significant predictors for both cats and kittens.

Machine learning

Machine learning is one possible tool that can be used to identify risk factors for animal adoption and predict the length of stay for animals in shelters. Machine learning is the ability to program computers to learn and improve all by itself using training experience [ 11 ]. The goal of machine learning is to develop a system to analyze big data, quickly deliver accurate and repeatable results, and to adapt to new data independently. A system can be trained to make accurate predictions by learning from examples of desired input-output data. More specifically, machine learning algorithms are utilized to detect classification and prediction patterns from large data and to develop models to predict future outcomes [ 12 ]. These patterns show the relationship between the attribute variables (input) and target variables (output) [ 13 ].

Widely used data mining tasks include supervised learning, unsupervised learning, and reinforcement learning [ 14 ]. Unsupervised learning involves the use of unlabeled datasets to train a system for finding hidden patterns within the data [ 15 ]. Clustering is an example of unsupervised learning. Reinforcement learning is where a system is trained through direct interaction with the environment by trial and error [ 15 ]. Supervised learning encompasses classification and prediction using labeled datasets [ 15 ]. These classification and regression algorithms are used to classify the output variable with a discrete label or predict the outcome as a continuous or numerical value. Traditional algorithms such as neural networks, decision trees, and logistic regression typically use supervised learning. Figure  1 provides a pictorial of the steps for developing and testing a predictive model.

Pictorial Representation of Developing a Predictive Model

Contributions to the literature

Although prior studies have investigated the impact of several factors, such as age and gender, on the length of stay, they focus on a single shelter, rather than multiple organizations, as in this study. The goal of this study is to investigate the length of stay of animals at shelters and the factors influencing the rate of animal adoption. The overall goal is to increase adoption rates of pets in animal shelters by utilizing several factors to predict the length of stay. Machine learning algorithms are used to predict the length of stay of each animal based on numerous factors (such as breed, size, and color). We address several objectives in this study that are listed below.

Identify risk factors associated with adoption rate and length of stay

Utilize the identified risk factors from collected data to develop predictive models

Compare statistical models to determine the best model for length of stay prediction

Exploratory Data results

From Fig.  2 , it is evident that the return of dogs is the highest outcome type at 43.3%, while Fig.  3 shows that the adoption of cats is the highest outcome type at 46.1%. Both figures illustrate that the euthanization of both cats and dogs is still prevalent (~ 20%). The results from Table 1 demonstrate that the longest time spent in the shelter is at 355 days by a male cat that is adopted and a female dog that is euthanized. Observing the results, adoption has the lowest variance among all animal types compared to the other outcome types. Adopted male cats have the lowest variance for days spent in the shelter, followed by female dogs. Female cats that are returned have the highest variance for days spent in the shelter.

Distribution of Outcome Types for Dogs

Distribution of Outcome Types for Cats

Figure  4 shows a comparison of cats and dogs for the three different outcome types. It is observed from the data that there are more dogs returned than cats. From Fig.  5 , it is observed that the number of days a dog stays in the shelter decreases as the age increases. This is not expected, as it is predicted that the number of days in a shelter would be lower for younger dogs and puppies. This observation could be due to having more data points for younger dogs.

Comparison of Outcome Types for Cats and Dogs

Age vs. Days in Shelter for Cats and Dogs

Machine learning results

Examining Table 2 , it is clear that the most proficient predictive model is developed by the gradient boosting algorithm for this dataset, followed by the random forest algorithm. The logistic regression algorithm appears to perform the worst with low precision, recall, and F1 score performance metrics for all categories of length of stay. For the prediction of low length of stay in a shelter, the random forest algorithm is the best performing model in comparison to the others at around 64–70% performance for precision, recall, and F1 score. The ANN algorithm is found to be the best when evaluating the precision and F1 score for medium length of stay, while the random forest algorithm is better for assessing recall. However, the performance of these models in predicting the medium length of stay for the given dataset is low for all three-performance metrics. The gradient boosting algorithm performs the best when predicting the high length of stay. Finally, the gradient boosting and random forest algorithms perform well when predicting the very high length of stay at around 70–80%.

Results from Table 2 also demonstrate that the model developed from the gradient boosting algorithm has a higher performance when predicting the high length of stay that leads to adoption, and when the outcome is euthanization. Evaluating the average of all three-performance metrics for all algorithms, the gradient boosting is the most proficient model at almost 60%, while logistic regression appears to be the worst. Table 2 also provides the computational time for each machine learning algorithm. For the given dataset, logistic regression runs the fastest at 9.41 s, followed by gradient boosting, artificial neural network, and finally, random forest running the longest. The gap in the performance measure ( pm ) is calculated by \( \frac{p{m}_{best}-p{m}_{worst}}{p{m}_{best}} \) , and is nearly 34, 39, and 32% for precision, recall, and F1 score, respectively.

Table 3 provides information on the top features or factors from each machine learning algorithm. Observing the table, we find that age (senior, super senior, and puppy), size (large and small), and color (multicolor) has a significant impact or influence on the length of stay. Specifically, we observe that older-aged animals (senior and/or super senior) appear as a significant factor for every algorithm. For the artificial neural network, older age is the #2 and #3 predictor, and super senior is the #2 predictor for the gradient boosting algorithm. Large and small-sized animals are also observed to be important features, as both are shown as the #1 predictor in the gradient boosting and ANN algorithms. The results also demonstrate that gender, animal type, other colors besides multicolor, middle age, and medium-sized animals did not significantly impact the length of stay.

Results from our study provided information on what factors are significant in influencing length of stay. Brown et al. [ 10 ] conducted research that found that age, breed designation, coat color, and coat pattern influenced the length of stay for cats in animal shelters. Similar to these studies, observations from our study also suggest that age and color have a significant impact or influence on the length of stay.

Determining which algorithm will develop the best model for the given set of data is critical to predict the length of stay and minimize the chances of euthanization. The goal of predictive analytics is to develop a model that best approximates the true mapping function for the relationship between the input and output variables. To approximate this function, parametric or non-parametric algorithms can be used. Parametric algorithms simplify the unknown function to a known form. Non-parametric algorithms do not make assumptions about the structure of the mapping function, allowing free learning of any functional form. In this study, we utilize both parametric (logistic regression and artificial neural network) and non-parametric (random forest and gradient boosting) algorithms on the given data. Observing the results from Table 2 , the gradient boosting and random forest (non-parametric algorithms) perform the best on the dataset. It is observed from the results that using a non-parametric approach leads to a better approximation of the true mapping function for the given records. These results also support prior studies on parametric versus non-parametric methods. Neely et al. [ 16 ] detailed the theoretical superiority of non-parametric algorithms for detecting pharmacokinetic and pharmacodynamic subgroups in a study population. The author suggests this superiority comes from the lack of assumptions made about the distribution of parameter values in a dataset. Bissantz et al. [ 17 ] discussed a resampling algorithm that evaluates the deviations between parametric and non-parametric methods to be noise or systematic by comparing parametric models to a non-parametric “supermodel”. Results demonstrate the non-parametric model to be significantly better. The use of algorithms that do not approximate the true function of the relationship between input and output provides better performance results for this application as well.

Current literature also supports the use of ensemble methods to increase prediction accuracy and performance. Dietterich [ 18 ] discussed the ongoing research into developing good ensemble methods as well as the discovery that ensemble algorithms are often more accurate than individual algorithms that are used to create them. Pandey, and S, T [ 19 ]. conducted a study to compare the accuracy of ensemble methodology on predicting student academic performance as research has demonstrated better results for composite models over a single model. This study applied ensemble techniques on learning algorithms (AdaBoost, Random Forest, Rotation Forest, and Bagging). For our study with the given records, the results support this claim. Both the gradient boosting and random forest algorithms are ensemble algorithms and performed the best on the animal shelter data.

Results from Table 2 demonstrate the best performance of the gradient boosting and random forest algorithm when the length of stay was classified as very high or the animal was euthanized. This is beneficial as the models can predict long stays where the outcome is euthanasia. This can lead to shelters identifying at-risk animals and implementing methods and solutions to ensure their adoption. These potential methods are the second phase of this research study, which will involve relocating animals to shelters where they will more likely be adopted. This phase is discussed in the future directions section.

Studies have been conducted evaluating euthanasia-related stress on workers (e.g., [ 1 ]). In other words, overpopulation not only leads to euthanasia but can, in turn, cause mental and emotional problems for the workers. For instance, Reeve et al. [ 20 ] evaluated the strain related to euthanasia among animal workers. Results demonstrated that euthanasia related strain was prevalent, and an increase in substance abuse, job stress, work causing family conflict, complaints, and low job satisfaction was observed. Predicting the length of stay for animals will aid in them being more likely to be adopted and will lead to fewer animals being euthanized, adding value not only to animals finding a home but also less stress on the workers.

The approach developed in this paper could be beneficial not only to reduce euthanasia but also to reduce overcrowding in shelters operated in countries where euthanasia of healthy animals is illegal, and all animals must be housed in shelters until adoption (or natural death). It is essential to develop an information system for a collaborative animal shelter network in which the entities can coordinate with each other, exchanging information about the animal inventory. Another benefit of this study is that it investigates applying machine learning to the animal care domain. Previous studies have looked into what factors influence the length of stay; however, this study utilizes these factors in addition to classification algorithms to predict how long an animal will stay in the shelter. Moreover, the use of a prescriptive analytics approach is discussed in this paper, where the predictions made by the machine learning algorithms will be used along with a goal programming model to decide in what shelter is an animal most likely to be adopted.

Limitations of this study include lack of behavioral data, limited sample size, and the use of simple algorithms. The first limitation, lack of behavioral data of the animal during intake and outcome, would be beneficial to develop a more comprehensive model. Though behavioral problems are harder to solve, having data would provide insight into how long these animals with behavioral issues are staying in shelters and what the outcome is. Studies have shown that behavioral problems play a significant role in preventing bonding between owners and their animals and one of the most common reasons cited for animal surrender [ 21 , 22 ]. These behavioral problems can include poor manners, too much energy, aggression, and destruction of the household. Dogs surrendered to shelters because of behavioral issues have also been shown to be less likely to be adopted or rehomed, and the ones that are adopted are more likely to be returned [ 21 ]. Studies have also been conducted to evaluate the effect of the length of time on the behavior of dogs in rescue shelters [ 23 , 24 , 25 ]. Most of them concluded that environmental factors led to changes in the behavior of dogs and that a prolonged period in a shelter may lead to unattractive behavior of dogs to potential owners. Acquiring information on behavioral problems gives more information for the algorithm to learn when developing the predictive model. This allows more in-depth predictions to be made on how long an animal will stay in a shelter, which could also aid in adoption. This approach can be used to shorten the length of stay, which makes sure that healthy animals are not developing behavioral problems in the shelters. It is not only crucial for the animal to be adopted, but also that the adoption is a good fit between owner and pet. Shortening the length of stay would also lessen the chance that the animal will be returned by the adopter because of behavior. Having this information will also allow shelters to find other shelters close by where animals with behavioral issues are more likely to be adopted. To overcome this limitation of the lack of data on behavioral problems, behavioral issues will be used as a factor and will be specifically asked for when acquiring data from shelters.

Another limitation includes collecting more data from animal shelters across the United States, allowing for more representative data to be collected and inputted into these algorithms. However, this presents a challenge due to most shelters being underfunded and low on staff. Though we reached out to shelters, most replied that they lacked the resources and staff to provide the information needed. Future work would include applying for funding to provide a stipend to staff for their assistance in gathering the data from respective shelters. With more data, the algorithm has more information to learn on, which could improve the performance metrics of the predictive models developed. There may also be other factors that show to be significant as more data is collected.

Finally, the last limitation is the use of simpler algorithms. This study considers basic ML algorithms. Nevertheless, in recent years, there has been development in the ML field of more complex networks. For instance, Zhong et al. [ 26 ] proposed a novel reinforcement learning method to select neural blocks and develop deep learning networks. Results demonstrated high efficiency in comparison to most of the previous deep network search approaches. Though only four algorithms were considered, future work would investigate deep learning networks, as well as bagging algorithms. Using more complex algorithms could ensure that if intricate patterns in the data are present, the algorithm can learn them.

Future direction

Phase 2: goal programming approach for making relocation decisions.

Using the information gathered in this study, we can predict the type of animals that are being adopted the most in each region and during each season of the year. To accomplish this, we utilize a two-phase approach. The first phase was leveraging the machine learning algorithms to predict the length of stay of each animal based on numerous factors (such as breed, size, and color). Phase-2 involves determining the best shelter to transport adoptable animals to increase the adoption rates, based on several conflicting criteria. This criterion includes predicted length of stay from phase-1, the distance between where the animal is currently housed and the potential animal shelters, transportation costs, and transportation time. Therefore, our goal is to increase adoption rates of pets in animal shelters by utilizing several factors to predict the length of stay, as well as determine the optimal animal shelter location where the animal will have the least amount stay in a shelter and most likely be adopted.

After predicting the length of stay of an incoming animal that is currently housed in the shelter l ′ using the machine learning algorithms, the next phase is to evaluate the potential relocation options for that animal. This strategic decision is specifically essential if the length of stay of the animal at its current location is high/very high. Nevertheless, while making this relocation decision, it is also necessary to consider the cost of transporting the animal between the shelters. For instance, if a dog is brought into a shelter in Houston, Texas, and is estimated to have a high/very high length of stay. Suppose if the dog is predicted to have a low length of stay at New York City and a medium length of stay at Oklahoma City, then a tradeoff has to be made between the relocation cost and the adoption speed. The objectives, length of stay, and relocation costs are conflicting and have to be minimized. Phase-2 attempts to yield a compromise solution that establishes a trade-off between these two criteria.

Goal programming (GP) is a widely used approach to solve problems involving multiple conflicting criteria. Under this method, each objective function is assigned as a goal, and a target value is specified for the individual criterion [ 27 ]. These target numbers can be fulfilled by the model with certain deviations, while the objective of the GP model is to minimize these deviations. Pertaining to this study, the desired values for the length of stay and relocation cost is pre-specified in the model and can be fulfilled with deviations. The GP model attempts to minimize these deviations. Thus, this technique attempts to produce a solution that is as close as possible to the targets, and the model solutions are referred to as the “most preferred solution” by prior studies (e.g., [ 28 , 29 ]).

As mentioned earlier, the primary task to be completed using this phase-2 goal programming approach is the relocation decisions considering the adoption speed and the cost of transporting the animal from the current location.

Model notations

Goal programming model formulation, goal constraints.

Objective 1: Minimize the overall length of stay of the animal under consideration (Eq. 1 ).

Goal constraint for objective 1: The corresponding goal constraint of objective 2 is given using Equation [ 30 ].

Objective 2: Minimize the overall relocation cost for transporting the animal under consideration (Eq. 3 ).

Goal constraint for objective 2: The corresponding goal constraint of objective 2 is given using Equation [ 18 ].

Hard constraints

Equation [ 9 ] ensures that the animal can be assigned to only one shelter.

The animal can be accommodated in shelter l only if there are a shelter capacity and type for that particular animal size category, and this is guaranteed using constraint [ 31 ]. It is important to note that both y and s are input parameters , whereas l is the set of shelters.

Equation [ 21 ] sets an upper limit on the length of stay category if the shelter l is assigned as the destination location. This prevents relocating animals to a shelter that might potentially have a high or very high length of stay.

Similarly, Equation [ 32 ] sets an upper limit on the relocation cost, if the shelter l is assigned as the destination location. This prevents relocating animals to a very far location. The current shelter location, l ′ , that is hosting the animal is an input parameter.

Objective function

Since the current problem focuses on minimizing the expected length of stay and relocation cost, the objective function of the goal programming approach is to reduce the sum of the weighted positive deviations given in Equations ([ 18 , 30 ], as shown in Equation [ 6 ].

where w g is the weight assigned for each goal g .

It is necessary to scale the deviation (since the objectives have different magnitudes as well as units) to avoid a biased solution.

If the scaling factors are represented by f g for goal g , then the scaled objective function is given in Equation [ 14 ].

Using this goal programming approach, the potential relocation options are evaluated considering the length of stay from phase-1. This phase-2 goal programming approach is useful, especially if the length of stay of the animal at its current location is high/very high, and a trade-off has to be made between relocation cost and length of stay. Phase-2 acts as a recommendation tool for assisting administrators with relocation decisions.

Nearly 3–4 million animals are euthanized out of the 6–8 million animals that enter shelters annually. The overall objective of this study is to increase the adoption rates of animals entering shelters by using key factors found in the literature to predict the length of stay. The second phase determines the best shelter location to transport animals using the goal programming approach to make relocation decisions. To accomplish this objective, first, the data is acquired from online sources as well as from numerous shelters across the United States. Once the data is acquired and cleaned, predictive models are developed using logistic regression, artificial neural network, gradient boosting, and random forest. The performance of these models is determined using three performance metrics: precision, recall, and F1 score.

The results demonstrate that the gradient boosting algorithm performed the best overall, with the highest precision, recall, and F1 score. Followed closely in second is the random forest algorithm, then the artificial neural network, and then finally, the logistic regression algorithm is the worst performer. We also observed from the data that the gradient boosting performed better when predicting the high or very high length of stay. Further observing the results, it is found that age for dogs (e.g., puppy, super senior), multicolor, and large and small size are important predictor variables.

The findings from this study can be utilized to predict how long an animal will stay in a shelter, as well as minimize their length of stay and chance of euthanization by determining which shelter location will most likely lead to the adoption of that animal. For future studies, we will implement phase 2, which will determine the best shelter location to transport animals using the goal programming approach to make relocation decisions.

Data description

A literature review is conducted to determine the factors that might potentially influence the length of stay for animals in shelters. These factors include gender, breed, age, and several other variables that are listed in Table 4 . These features will be treated as input variables for the machine learning algorithms. Overall, there are eight input or predictor variables and one output variable, which is the length of stay.

Animal shelter intake and outcome data are publicly made available by several state/city governments on their website (e.g., [ 33 , 34 ]), specifically in several southern and south-western states. These online sources provide datasets for animal shelters from Kentucky (150,843 data rows), California (334,016), Texas (155,115), and Indiana (4132). Since there is no nationwide database for animal shelters, information is also collected through individual animal shelters that conduct euthanization of animals. We contacted over 100 animal shelters across the United States and inquired for data on the factors mentioned in Table 4 . We received responses from 20 of the animal shelters that were contacted. Most responses received stated there was not enough staff or resources to be able to provide this information. From the responses that were received back, only four shelters were able to provide any information. Of those four, only two of the datasets contained the factors and information needed, which are Colorado (8488 data rows) and Arizona (4, 667 data rows).

The data that is collected from the database and animal shelters included information such as animal type, intake and outcome date, gender, color, breed, and intake and outcome status (behavior of animal entering the shelter and behavior of animal at outcome type). These records also included information on several types of animals, such as dogs, cats, birds, rabbits, and lizards. For this study, the focus is on dogs and cats. After filtering through these records, we found that only California, Kentucky, Colorado, Arizona, and Indiana had all of the factors needed for the study. Upon downloading data from the database and receiving data from the animal shelters, the acquired data underwent data integration, data transformation, and data cleaning (as detailed in Fig.  1 ). After data pre-processing, there are over 113,000 animal records.

Data cleaning methods

Next, data cleaning methods are utilized to detect discrepancies in the data, such as missing values, erroneous data, and inconsistencies. Data cleaning is an essential step for obtaining unbiased results [ 35 , 36 ]. In other words, identifying and cleaning erroneous data must be performed before inputting the data into the algorithm as it can significantly impact the output results.

The following is a list of commonly used data cleaning techniques in the literature [ 11 ]:

Substitution with Median: Missing or incorrect data are replaced with the median value for that predictor variable.

Substitution with a Unique Value: Erroneous data are replaced with a value that does not fall within the range that the input variables can accept (e.g., a negative number)

Discard Variable and Substitute with a Median: When an input variable has a significant number of missing values, these values are removed from the dataset, and the features that remain with missing or erroneous values are replaced with the median.

Discard Variable and Substitute with a Unique Value: Input variables with a significant number of missing values are removed from the dataset, and the features that remain with missing or erroneous values are coded as − 1.

Remove Incomplete Rows Entirely: Incomplete Rows are removed from the dataset.

Data preprocessing

Some animal breeds are listed in multiple formats and are changed to maintain uniformity. An example of this is a Russian Blue cat, which is formatted in several ways such as “Russian”, “Russian Blue”, and “RUSSIAN BLUE”. Animals with multiple breeds such as “Shih Tzu/mix” or “Shih Tzu/Yorkshire Terr” are classified as the first breed listed. Other uncommon breeds are classified as “other” for simplicity. Finally, all animal breeds are summarized into three categories (small, medium, or large) using the American Kennel Clubs’ breed size classification [ 37 ]. Part of the data cleansing process also includes categorizing multiple colors found throughout the sample size into five distinct color categories (brown, black, blue, white, and multicolor). We classified age into five categories for dogs and cats (puppy or kitten, adolescent, adult, senior, super senior). The puppy or kitten category includes data points 0–1 year, adolescence includes data points 2–3 years old, adulthood includes animals 4–7 years of age, and senior animals are 8–10 years of age. Any animal that is older than ten years are categorized as a super senior, based on the recommendations provided in Wapiti Labs [ 38 ].

As mentioned previously, the output variable is the length of stay and is classified as low, medium, high, and very high/euthanization. The length of stay is calculated by taking the difference between the intake date and outcome date. To remove erroneous data entries and special cases, the number of days in the animal shelter is also capped at a year. The “low” category represents animals that are returned (in which case, they are assigned the days in the shelter as 0) or spent less than 8 days before getting adopted. It is important to keep these animals at the shelter so that the owner may find them or they are transferred to their new homes. Animals that stayed in a shelter for 9–42 days and are adopted are categorized as “medium” length of stay. The “high” category is given to animals that stayed in the shelter for 43–365 days. Finally, animals that are euthanized are categorized as “very high”.

After integrating all data points from each animal shelter, the sample size includes 119,691 records. After the evaluation of these data points, 5436 samples are found to have miscellaneous (such as a negative length of stay) or missing values. After applying data cleaning techniques, the final cleaned dataset includes 114,256 data points, with 50,466 cat- and 63,790 dog-records.

Machine learning algorithms to predict the length of stay

The preprocessed records are then separated into training and testing datasets based on the type of classification algorithm used. Studies have demonstrated the need for testing and comparing machine learning algorithms, as the performance of the models depends on the application. While an algorithm may develop a predictive model that performs well in one application, it may not be the best performing model for another. A comparison between the statistical models is conducted to determine the overall best performing model. In this section, we provide a description as well as the advantages of each classification algorithm that is utilized in this study.

Logistic regression

Logistic regression (LR) is a machine learning algorithm that is used to understand the probability of the occurrence of an event [ 39 ]. It is typically used when the model output variable is binary or categorical (see Fig.  6 ), unlike linear regression, where the dependent variable is numeric [ 40 ]. Logistic regression involves the use of a logistic function, referred to as a “sigmoid function” that takes a real-valued number and maps it into a value between 0 and 1 [ 41 ]. The probability that the length of stay of the animal at a specific location will be low, medium, high, or very high, is computed using the input features discussed in Table 4 .

Pictorial Representation of the Logistic Regression Algorithm

The linear predictor function to predict the probability that the animal in record i has a low, medium, high, and very high length of stay categories is given by Equations ( 11 ) –[ 3 ], respectively.

Where β v , l is a set of multinomial logistic regression coefficients for variable v of the length of stay category l , and x v , i is the input feature v corresponding to data observation i .

Artificial neural network

Artificial Neural Network (ANN) algorithms were inspired by the brain’s neuron, which transmits signals to other nerve cells [ 40 , 42 ]. ANN’s were designed to replicate the way humans learn and were developed to imitate the operational sequence in which the body sends signals in the nervous system [ 43 ]. In an ANN, there exists a network structure with directional links connecting multiple nodes or “artificial neurons”. These neurons are information-processing units, and the ties that connect them represent the relationship between each of the connected neurons. Each ANN consists of three layers - the input layer, hidden layer, and the output layer [ 32 , 44 ]. The input layer is where each of the input variables is fed into the artificial neuron. The neuron will first calculate the sum of multiple inputs from the independent variables. Each of the connecting links (synapses) from these inputs has a characterized weight or strength that has a negative or positive value [ 32 ]. When new data is received, the synaptic weight changes, and learning will occur. The hidden layer learns the relationship between the input and output variables, and a threshold value determines whether the artificial neuron will fire or pass the learned information to the output layer, as shown in Fig.  7 . Finally, the output layer is where labels are given to the output value, and backpropagation is used to correct any errors.

Pictorial Representation of the Artificial Neural Networks

Random Forest

The Random Forest (RF) algorithm is a type of ensemble methodology that combines the results of multiple decision trees to create a new predictive model that is less likely to misclassify new data [ 30 , 45 ]. Decision Trees have a root node at the top of the tree that consists of the attribute that best classifies the training data. The attribute with the highest information gain (given in Eq. 16 ) is used to determine the best attribute at each level/node. The root node will be split into more subnodes, which are categorized as a decision node or leaf node. A decision node can be divided into further subnodes, while a leaf node cannot be split further and will provide the final classification or discrete label. RF algorithm uses mtree and ntry as the two main parameters in developing the multiple parallel decision trees. Mtree specifies how many trees to train in parallel, while ntry defines the number of independent variables or attributes to choose to split each node [ 30 ].. The majority voting from all parallel trees gives the final prediction, as given in Fig.  8 .

Pictorial Representation of the Random Forest Algorithm

Gradient boosting

Boosting is another type of ensemble method that combines the results from multiple predictive algorithms to develop a new model. While the RF approach is built solely on decision trees, boosting algorithms can use various algorithms such as decision trees, logistic regression, and neural networks. The primary goal of boosting algorithms is to convert weak learners into stronger ones by leveraging weighted averages to identify “weak classifiers” [ 31 ]. Samples are assigned an initial uniformed weight, and when incorrectly labeled by the algorithm, a penalty of an increase in weight is given [ 46 ]. On the other hand, samples that are correctly classified by the algorithm will decrease in weight. This process of re-weighing is done until a weighted vote of weak classifiers is combined into a robust classifier that determines the final labels or classification [ 46 ]. For our study, gradient boosting (GB) will be used on decision trees for the given dataset, as illustrated in Fig.  9 .

Pictorial Representation of Boosting Algorithm

Machine learning model parameters

The clean animal shelter data is split into two datasets: training and testing data. These records are randomly placed in the two groups to train the algorithms and to test the model developed by the algorithm. 80% of the data is used to train the algorithm, while the other 20% is used to test the predictive model. To avoid overfitting, a tenfold cross-validation procedure is used on the training data. There are no parameters associated with the machine learning of logistic regression algorithms. However, a grid search method is used to tune the parameters of the random forest, gradient boosting, and artificial neural network algorithms. This allows the best parameter in a specific set to be chosen by running an in-depth search by the user during the training period.

The number of trees in the random forest and gradient boosting algorithms is changed from 100 to 1000 in increments of 100. A learning rate of 0.01, 0.05, and 0.10 is used based on the recommendations of previous studies [ 47 ]. The minimum observations for the trees’ terminal node are set to vary from 2 to 10 in increments of one, while the splitting of trees varies from 2 to 10 in increments of two. A feed-forward method is used to develop the predictive model using the artificial neural network algorithm. The feed-forward algorithm consists of three layers (input, hidden, output) as well as backpropagation learning. The independent and dependent variables represent the input and output layers. Since the input and output layers are already known, an optimal point is reached for the number of nodes when between 1 and the number of predictors. This means that for our study, the nodes of the hidden layer vary from 1 to 8. The learning rate values used to train the ANN are 0.01, 0.05, and 0.10.

To find the optimal setting for each machine learning algorithm, a thorough search of their corresponding parameter space is performed.

Performance measures

In this study, we use three performance measures to evaluate the ability of machine learning algorithms in developing the best predictive model for the intended application. The measures considered are precision, F1 score, and sensitivity/recall to determine the best model given the inputted data samples. Table 5 provides a confusion matrix to define the terms used for all possible outcomes.

Precision evaluates the number of correct, true positive predictions by the algorithm while still considering the incorrectly predicted positive when it should have been negative (Eq. 17 ). By having high precision, this means that there is a low rate of false positives or type I error. Sensitivity or recall evaluates the number of true positives that are correctly predicted by the algorithm while considering the incorrectly predicted negative when it should have been positive (Eq. 18 ). Recall is a good tool to use when the focus is on minimizing false negatives (type II error). F1 score (shown in Eq. 19 ) evaluates both type I and type II errors and assesses the ability of the model to resist false positives and false negatives. This performance metric evaluates the robustness (low number of missed classifications), as well as the number of data points that are classified correctly by the model.

Availability of data and materials

Most of the datasets used and/or analyzed during the current study were publicly available online as open source data. The data were available in the website details given below:

https://data.bloomington.in.gov/dataset

https://data.louisvilleky.gov/dataset

https://data.sonomacounty.ca.gov/Government

We also obtained data from Sun Cities 4 Paws Rescue, Inc., and the Rifle Animal Shelter. No administrative permission was required to access the raw data from these shelters.

Abbreviations

Logistic Regression

Artificial Neural Network

Gradient Boosting

Goal Programming

Coefficient of Variation

Anderson KA, Brandt JC, Lord LK, Miles EA. Euthanasia in animal shelters: Management's perspective on staff reactions and support programs. Anthrozoös. 2013;26(4):569–78. https://doi.org/10.2752/175303713X13795775536057 .

Article   Google Scholar  

Clevenger J, Kass PH. Determinants of adoption and euthanasia of shelter dogs spayed or neutered in the University of California veterinary student surgery program compared to other shelter dogs. J Veterinary Med Educs. 2003;30(4):372–8.

Animal Humane Society. (n.d.). Retrieved November 2019, from https://www.animalhumanesociety.org/ .

Home. (2016, July 15). Retrieved November 2019, from http://www.americanhumane.org/ .

Rogelberg SG, DiGiacomo N, Reeve CL, Spitzmüller C, Clark OL, Teeter L, et al. What shelters can do about euthanasia-related stress: an examination of recommendations from those on the front line. J Appl Anim Welf Sci. 2007;10(4):331–47. https://doi.org/10.1080/10888700701353865 .

Article   CAS   PubMed   Google Scholar  

Hennessy MB, Voith VL, Mazzei SJ, Buttram J, Miller DD, Linden F. Behavior and cortisol levels of dogs in a public animal shelter, and an exploration of the ability of these measures to predict problem behavior after adoption. Appl Anim Behav Sci. 2001;73(3):217–33.

Shore ER. Returning a recently adopted companion animal: Adopters' reasons for and reactions to the failed adoption experience. J Appl Anim Welf Sci. 2005;8(3):187–98.

Article   CAS   Google Scholar  

Morris KN, Gies DL. Trends in intake and outcome Data for animal shelters in a large U.S. metropolitan area, 1989 to 2010. J Appl Anim Welf Sci. 2014;17(1):59–72. https://doi.org/10.1080/10888705.2014.856250 .

Fantuzzi JM, Miller KA, Weiss E. Factors relevant to adoption of cats in an animal shelter. J Appl Anim Welf Sci. 2010;13(2):174–9.

Brown WP, Morgan KT. Age, breed designation, coat color, and coat pattern influenced the length of stay of cats at a no-kill shelter. J Appl Anim Welf Sci. 2015;18(2):169–80.

Srinivas, S., & Rajendran, S. (2017). A Data-driven approach for multiobjective loan portfolio optimization using machine-learning algorithms and mathematical programming. In big Data analytics using multiple criteria decision-making models (pp. 175-210): CRC press.

Waller MA, Fawcett SE. Data science, predictive analytics, and big Data: a revolution that will transform supply chain design and management. J Bus Logist. 2013;34(2):77–84.

Kantardzic M. DATA MINING: concepts, models, methods, and algorithms. 2nd ed: IEEE: Wiley; 2019.

Jordan MI, Mitchell TM. Machine learning: trends, perspectives, and prospects. Science. 2015;349(6245):255–60.

Kavakiotis I, Tsave O, Salifoglou A, Maglaveras N, Vlahavas I, Chouvarda I. Machine learning and Data mining methods in diabetes research. Computational Structural Biotechnol J. 2017;15:104–16. https://doi.org/10.1016/j.csbj.2016.12.005 .

Neely MN, van Guilder MG, Yamada WM, Schumitzky A, Jelliffe RW. Accurate detection of outliers and subpopulations with Pmetrics, a nonparametric and parametric pharmacometric modeling and simulation package for R. Ther Drug Monit. 2012;34(4):467–76. https://doi.org/10.1097/FTD.0b013e31825c4ba6 .

Article   PubMed   PubMed Central   Google Scholar  

Bissantz N, Munk A, Scholz A. Parametric versus non-parametric modelling? Statistical evidence based on P-value curves. Mon Not R Astron Soc. 2003;340(4):1190–8. https://doi.org/10.1046/j.1365-8711.2003.06377.x .

Dietterich TG. Ensemble methods in machine learning. Berlin: Heidelberg; 2000.

Book   Google Scholar  

Pandey M, S, T. A comparative study of ensemble methods for students' performance modeling. Int J Computer ApplS. 2014;103:26–32. https://doi.org/10.5120/18095-9151 .

Reeve CL, Rogelberg SG, Spitzmüller C, Digiacomo N. The caring-killing paradox: euthanasia-related strain among animal-shelter Workers1. J Appl Soc Psychol. 2005;35(1):119–43. https://doi.org/10.1111/j.1559-1816.2005.tb02096.x .

Gates MC, Zito S, Thomas J, Dale A. Post-adoption problem Behaviours in adolescent and adult dogs rehomed through a New Zealand animal shelter. Animals : an open access journal from MDPI. 2018;8(6):93. https://doi.org/10.3390/ani8060093 .

Weiss E, Gramann S, Drain N, Dolan E, Slater M. Modification of the feline-Ality™ assessment and the ability to predict adopted Cats' behaviors in their new homes. Animals : an open access journal from MDPI. 2015;5(1):71–88. https://doi.org/10.3390/ani5010071 .

Normando S, Stefanini C, Meers L, Adamelli S, Coultis D, Bono G. Some factors influencing adoption of sheltered dogs. Anthrozoös. 2006;19(3):211–24.

Protopopova A, Mehrkam LR, Boggess MM, Wynne CDL. In-kennel behavior predicts length of stay in shelter dogs. PLoS One. 2014;9(12):e114319.

Wells DL, Graham L, Hepper PG. The influence of length of time in a rescue shelter on the behaviour of Kennelled dogs. Anim Welf. 2002;11(3):317–25.

CAS   Google Scholar  

Zhong G, Jiao W, Gao W, Huang K. Automatic design of deep networks with neural blocks. Cogn Comput. 2020;12(1):1–12.

Rajendran S, Ravindran AR. Multi-criteria approach for platelet inventory management in hospitals. Int J Operational ResS. 2020;38(1):49–69.

Bastian ND, McMurry P, Fulton LV, Griffin PM, Cui S, Hanson T, Srinivas S. The AMEDD uses goal programming to optimize workforce planning decisions. Interfaces. 2015;45(4):305–24.

Rajendran S, Ansaripour A, Kris Srinivasan M, Chandra MJ. Stochastic goal programming approach to determine the side effects to be labeled on pharmaceutical drugs. IISE Transactions on Healthcare Systems Engineering. 2019;9(1):83–94.

Cutler DR, Edwards TC Jr, Beard KH, Cutler A, Hess KT, Gibson J, Lawler JJ. Random forests for classification in ECOLOGY. Ecology. 2007;88(11):2783–92.

Friedman J, Hastie T, Tibshirani R. Additive logistic regression: a statistical view of boosting (with discussion and a rejoinder by the authors). Ann Stat. 2000;28(2):337–407.

Ge Z, Song Z, Ding SX, Huang B. Data mining and analytics in the process industry: the role of machine learning. IEEE Access. 2017;5:20590–616.

Open Data: City of Austin Texas: Open Data: City of Austin Texas. (n.d.). Retrieved March 2019, from https://data.austintexas.gov//Health-and-Community-Services/Austin-Animal-Center-Outcomes/9t4d-g238 .

County of Sonoma: Open Data: Open Data. (n.d.). Retrieved March 2019, from https://data.sonomacounty.ca.gov/Government/Animal-Shelter-Intake-and-Outcome/924a-vesw .

Kambli A, Sinha AA, Srinivas S. Improving campus dining operations using capacity and queue management: a simulation-based case study. J Hosp Tour Manag. 2020;43:62–70.

Rajendran S, Zack J. Insights on strategic air taxi network infrastructure locations using an iterative constrained clustering approach. Transport Res Part E: Logistics and Transportation Review. 2019;128:470–505.

American Kennel Club. (n.d.). Retrieved November 2019, from http://www.akc.org/ .

Elk Antler Supplements & Chews: Wapiti Labs, Inc. (n.d.). Retrieved November 2019, from https://www.wapitilabsinc.com/ .

Bursac Z, Gauss CH, Williams DK, Hosmer DW. Purposeful selection of variables in logistic regression. Source Code for Biol Med. 2008;3(1):17.

Delen D, Walker G, Kadam A. Predicting breast cancer survivability: a comparison of three data mining methods. Artif Intell Med. 2005;34(2):113–27.

Kim A, Song Y, Kim M, Lee K, Cheon JH. Logistic regression model training based on the approximate homomorphic encryption. BMC Med Genet. 2018;11(4):83.

Google Scholar  

Srinivas S, Ravindran AR. Optimizing outpatient appointment system using machine learning algorithms and scheduling rules: a prescriptive analytics framework. Expert Syst Appl. 2018;102:245–61. https://doi.org/10.1016/j.eswa.2018.02.022 .

LeCun Y, Bengio Y, Hinton G. Deep learning. Nature. 2015;521:436.

Shih H, Rajendran S. Comparison of time series methods and machine learning algorithms for forecasting Taiwan blood Services Foundation’s blood supply. Journal of healthcare engineering. 2019;2019.

Srinivas S, Salah H. Consultation length and no-show prediction for improving appointment scheduling efficiency at a cardiology clinic: a data analytics approach. Int J Med Inform. 2020;145:104290.

Rokach L. Ensemble-based classifiers. Artif Intell Rev. 2010;33(1):1–39.

Srinivas S. A machine learning-based approach for predicting patient punctuality in ambulatory care centers. Int J Environ Res Public Health. 2020;17(10):3703.

Download references

Acknowledgments

We would like to thank the Sun Cities 4 Paws Rescue, Inc., and the Rifle Animal Shelter for providing the length of stay reports in order to complete this study.

This research was not funded by any agency/grant.

Author information

Authors and affiliations.

Department of Bioengineering, University of Missouri Columbia, Columbia, MO, 65211, USA

Janae Bradley

Department of Industrial and Manufacturing Systems Engineering, University of Missouri Columbia, Columbia, MO, 65211, USA

Suchithra Rajendran

Department of Marketing, University of Missouri Columbia, Columbia, MO, 65211, USA

You can also search for this author in PubMed   Google Scholar

Contributions

JB performed data mining, data cleaning and analyses of the animal shelter data and machine learning algorithms. JB was also a major contributor in writing the manuscript. SR performed data mining, cleaning, and analyses of the machine learning algorithms, as well as the goal programming. All authors have read and approved the final manuscript.

Corresponding author

Correspondence to Suchithra Rajendran .

Ethics declarations

Ethics approval and consent to participate.

Most of the datasets used in this study are open source and are publicly available. The remaining data was collected from animal shelters with their consent to use the data for research purposes.

Consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests.

Additional information

Publisher’s note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ . The Creative Commons Public Domain Dedication waiver ( http://creativecommons.org/publicdomain/zero/1.0/ ) applies to the data made available in this article, unless otherwise stated in a credit line to the data.

Reprints and permissions

About this article

Cite this article.

Bradley, J., Rajendran, S. Increasing adoption rates at animal shelters: a two-phase approach to predict length of stay and optimal shelter allocation. BMC Vet Res 17 , 70 (2021). https://doi.org/10.1186/s12917-020-02728-2

Download citation

Received : 07 January 2020

Accepted : 22 December 2020

Published : 05 February 2021

DOI : https://doi.org/10.1186/s12917-020-02728-2

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Animal shelter
• High euthanization rates
• Machine learning algorithms
• Prediction models
• Goal programming approach
• Decision support tool

BMC Veterinary Research

ISSN: 1746-6148

• General enquiries: [email protected]

Welcome to JSAR – Free, peer-reviewed research

The Journal of Search and Rescue (JSAR) is the world’s only free peer-reviewed electronic journal for the collation and distribution of original scholarly material on search and rescue.

JSAR serves those with a direct interest in all disciplines of search and rescue (SAR) including: rope rescue, water (flat, swift and marine), ice rescue, wilderness search and rescue, structural collapse rescue, trench collapse rescue, cave rescue, dive rescue, motor vehicle extrication, canine search, technical animal rescue, air rescue and mines rescue.

The journal accepts article submissions from these and other SAR disciplines.

You can read and download all published articles on our Issues  page

We welcome questions, comments, suggestions and offers to help develop JSAR. Contact us at [email protected] .

Ethical concerns in rescue robotics: a scoping review

• Review Paper
• Open access
• Published: 02 July 2021
• Volume 23 , pages 863–875, ( 2021 )

Cite this article

You have full access to this open access article

• Linda Battistuzzi   ORCID: orcid.org/0000-0003-2340-3059 1 ,
• Carmine Tommaso Recchiuto 1 &
• Antonio Sgorbissa 1  

9083 Accesses

9 Citations

Explore all metrics

Rescue operations taking place in disaster settings can be fraught with ethical challenges. Further ethical challenges will likely be introduced by the use of robots, which are expected to soon become commonplace in search and rescue missions and disaster recovery efforts. To help focus timely reflection on the ethical considerations associated with the deployment of rescue robots, we have conducted a scoping review exploring the relevant academic literature following a widely recognized scoping review framework. Of the 429 papers identified by the first screening, six fulfilled the selection criteria of our literature review. Quantitative data synthesis showed that a subset of the papers includes a qualitative experimental exploration of the ethical issues at hand, with workshops involving both experts and potential users. Most use simulations or scenarios to anticipate the ethical implications and other consequences of using robots in search and rescue missions. Qualitative text analysis identified seven core ethically relevant themes: fairness and discrimination; false or excessive expectations; labor replacement; privacy; responsibility; safety; trust. Our results suggest that the literature on ethics in rescue robotics is scant and disparate, but the papers identified uniformly endorsed a proactive approach to handling the ethical concerns associated with the use of robots in disaster scenarios.

Similar content being viewed by others

Exploring the Ethical Landscape of Robot-Assisted Search and Rescue

Ethical Considerations Associated with “Humanitarian Drones”: A Scoping Literature Review

Ning Wang, Markus Christen & Matthew Hunt

Ethics, War, and Robots

Avoid common mistakes on your manuscript.

Rescue robotics is a relatively young discipline within field robotics. Its goal is to provide rescuers in operation areas with the ability to sense and act at a distance from the site of disasters (Murphy, 2014 ), i.e. “phenomena caused by environmental or man-made events that lead to fatalities, injuries, stress, physical damage and economic breakdown of great significance.” (Cuny, 1992 ).

Rescue robots can enable operators to access areas in harsh conditions that would be inaccessible, or too dangerous or slow for humans to enter. They can also serve as remote sensing platforms that make it possible for humans to interact with the destroyed environment (Adams et al., 2014 ; Kochersberger et al., 2014 ; Stefanov & Evans, 2014 ). A rescue robot can, for example, help visually examine and map the interior of a collapsed building, inspect damage (Devault, 2000 ; Ellenberg et al., 2015 ; Lattanzi & Miller, 2017 ; Recchiuto & Sgorbissa, 2017 ; Torok et al., 2014 ), place acoustic, or thermal, or seismic sensors to monitor the situation, or quickly remove heavy rubble to facilitate extricating victims (Murphy & Stover, 2007 ; Murphy et al., 2009 ; Steimle et al., 2009 ). Providing this kind of rapid access and intervention should translate into fewer lives lost, lesser injuries and, overall, faster recovery from the disaster itself (Murphy, 2014 ).

The first reported use of rescue robots at a disaster site dates to 2001, when the Center for Robot-Assisted Search and Rescue used robots from the DARPA Tactical Mobile Robots program at the World Trade Center disaster in New York City (Murphy, 2014 ). Since then, rescue robots have been deployed across the world in mine accidents, earthquakes, mudslides, nuclear disasters, hurricanes, oil spills and building collapses, thereby gaining widespread public prominence. In the coming years, owing to the growing impact of natural and man-made disasters, the need for such robots is expected to increase across all phases of the disaster life-cycle (Murphy, 2014 ). In light of this broader view of their role, “rescue robots” are often termed “disaster robots”; the two terms will thus be used interchangeably throughout this paper, as will the terms “rescuer”, “responder” and “operator”.

The types of robots that are employed in disasters include Unmanned Ground Vehicles (UGV), which carry a range of sensors and are typically equipped with tracks to traverse unstructured terrains, Unmanned Aerial Vehicles (UAV), which can provide aerial support for disaster response operations, and Unmanned Marine Vehicles (UMV), which can, for instance, carry out underwater inspections and insert mitigation devices. Although most of these robots are controlled by humans, semi-autonomous systems that reduce the need for low-level control by operators are becoming more frequent (Birk & Carpin, 2006 ; Delmerico et al., 2019 ; Zuzanek et al., 2014 ).

Operations taking place in disaster settings are fraught with ethical challenges. Many of those challenges are associated with the hazardous, chaotic, and pressure-filled conditions under which responders must operate and the lack of time, materials, and capacity that characterizes their work. Choices regarding, for instance, where to concentrate rescue efforts, what kind of risks should be taken, whom to search for first, who should be given priority treatment, who must be left to wait, and how to make optimal use of the limited resources available, are morally burdensome (Gustavsson et al., 2020 ), and the consequences of those choices can weigh on victims and responders, but also on other stakeholders.

Policies and guidelines exist to support responders in their work (Medical Ethics Manual. World Medical Association., 2015 ; The ICN Code of Ethics for Nurses. International Council of Nurses., 2012 ), (Green et al., 2003 ). Limited guidance is available, however, for those who do not have medical roles and for ethically informed, practical decision-making in specific disaster settings (Gustavsson et al., 2020 ). This situation is exacerbated by the generalized lack of specific training programs to develop the knowledge and skills required for such decision-making (Gustavsson et al., 2020 ). Rescue robots’ increasing presence in operation areas is likely to carry an additional layer of ethical complexity, which will depend in part on what type of robots are used and the specific contexts in which they are deployed.

In some domains, e.g. in the industry, military and in education, ethical concerns regarding the application of robots have received much attention (Lichoki et al., 2011 ). A great deal of reflection has also been devoted to the use of robots in healthcare, looking at their impact on the privacy (Sharkey & Sharkey, 2012 ), human rights (Sharkey & Sharkey, 2011 ), and autonomy of patients (Sparrow, 2016 ). However, not much effort seems to have been dedicated to exposing and elucidating the ethical issues that may emerge when robots are used in disaster settings (Harbers et al., 2017 ). Therefore, to help focus timely ethical reflection on rescue robotics before rescue robots become commonplace and their use ubiquitous, we have conducted a scoping review of the relevant literature.

We followed the well-known scoping review framework by Arksey and O’Malley (Arksey & O’Malley, 2005 ) and subsequent recommendations (Colquhoun et al., 2014 ; Levac et al., 2010 ) for conducting and reporting scoping reviews.

Although there is no clear consensus on their definition or purpose, scoping reviews are commonly described as tools to map or synthesize a range of evidence in order to convey the size and scope of a research field (Levac et al., 2010 ).

According to Arksey and O’Malley, by conducting a scoping study researchers can survey the extent, range, and nature of research activity in a given field, establish whether a full systematic review is warranted, summarize and disseminate research evidence, or identify gaps in the existing literature (Arksey & O’Malley, 2005 ). Unlike systematic reviews, scoping studies do not typically provide an assessment of the quality of the studies covered (Grant & Booth, 2009 ; Rumrill et al., 2010 ). Unlike narrative or literature reviews, they require analytical reinterpretation of the literature. Scoping reviews are particularly relevant when the literature on a topic is complex or heterogeneous, can cover findings from a range of different study designs and methods, and may be especially useful when evidence on a topic is emerging (Levac et al., 2010 ). Thus, they are well-suited to synthesizing the literature on a topic that has yet to be comprehensively mapped, such as the one we have endeavored to review here.

This scoping review surveyed published literature and ethics approval was not required.

We identified papers discussing ethical issues associated with the use of rescue robots using a two-tiered search strategy: (1) searching five databases (Google Scholar, IEEE Xplore, Science Direct, Scopus, and the Web of Science), and (2) searching the references in the documents that were selected for the final qualitative synthesis. We searched title, abstract and keywords for the terms: ethics AND (“rescue robot” OR “disaster robot”). Query logic was modified to adapt to the language used by each engine or database. Only the first 250 hits retrieved in Google Scholar ordered by relevance were considered, in accordance with the methods used in numerous similar reviews. The search initially yielded 429 entries. Following the recommendations by Pham and colleagues (Pham et al., 2014 ), the subsequent study selection process was conducted using the Preferred Reporting Items for Systematic Reviews and Meta-Analyses ( http://prisma-statement.org ) as a guide (see Fig.  1 ).

PRISMA screening process for identified papers

At the first screening stage, one researcher reviewed the titles and abstracts of papers. Only papers written in English were included. Duplicates, papers not related to the topic, as well as theses, articles from the popular press, reports, non-reviewed books and book chapters, presentations and opinion pieces were excluded. This gave 42 papers which underwent further screening. The second eligibility screening was conducted independently by two members of the research team. Each researcher evaluated the papers against the inclusion and exclusion criteria and independent results were compared. A third researcher was involved to resolve any disagreements about paper eligibility.

Documents were included if they fulfilled all the following criteria:

Papers published in a peer-reviewed outlet contained in Google Scholar, IEEE Xplore, Science Direct, Scopus, or the Web of Science;

Papers in English;

Papers that included the relevant search terms as previously defined;

Papers in which ethical issues associated with rescue robots were the main focus or at least were addressed in their own part or section;

Papers that were published in 2001 or later.

In addition, the following items comprised our exclusion criteria:

Papers that focused on ethical issues in robotics but only casually mentioned rescue robots;

Papers that focused on rescue robotics but only casually mentioned any associated ethical issues.

Once the papers that fulfilled all of the criteria were identified, each of these papers’ reference list was screened for additional relevant documents.

Based on the recommendations by Levac and colleagues (Levac et al., 2010 ), we performed a descriptive quantitative synthesis and a thematic analysis. Following Arksey and O’Malley (Arksey & O’Malley, 2005 ), our descriptive quantitative summary includes the details of the articles identified, year of publication, discipline/field of inquiry, type of research, type of robot investigated etc. For the thematic analysis, the papers were coded following a multi-step process including open coding, axial coding and selective coding, using the Dedoose web-based application ( www.dedoose.com ). In the first phase, units of meaning were identified and labeled to allow categories to emerge from the data (open coding). Open codes were then categorized, with similar codes grouped, refined and combined into larger themes (axial coding). The conceptually stable thematic patterns that emerged were then organized and grouped into higher order conceptual themes Selective coding involved the integration and refinement of these concepts (Corbin & Strauss, 2008 ). Finally, findings were integrated and validated through discussion among all members of the research team.

Quantitative data synthesis

Six papers fulfilled the selection criteria of our literature review. Most were published in scientific journals (4/6) and all were distributed across different outlets (see Table 1 ). No relevant publications were identified in any of the papers’ list of references.

None of the papers identified were published before 2008, and 5/6 were published between 2014 and 2020 (see Table 1 ).

The studies are mostly situated in robotics (n = 3), robot ethics (n = 3) and machine ethics (n = 3), but also in technology assessment (n = 2), and information systems (n = 1). Most of the papers include elements from different disciplines (see Table 2 ).

Most studies are conceptual and/or technological but 2/6 feature a mixed approach including an experimental exploration of the ethical issues at hand. UGVs are discussed in most of the publications (4/6), with the remaining focusing on UAVs or both UGVs and UAVs (see Table 2 ).

Five of the six papers refer to a method of practicing ethics in Research and Innovation, and four use simulations or scenarios to anticipate the ethical implications and other consequences of using robots in search and rescue missions; one instead considers two case studies of robot system deployments in search and rescue settings to test socio-ethical approaches to the development of robots (see Table 2 ).

Qualitative text analysis

Qualitative text analysis highlighted seven core ethical concerns: fairness and discrimination; false or excessive expectations; labor replacement; privacy; responsibility; safety; trust. Reported in Table 2 are the occurrences of these themes across the six publications.

Two of the papers included discussion of ethical concerns that emerged from an inductive process involving qualitative research with stakeholders (Carlsen et al., 2015 ; Harbers et al., 2017 ). Three papers, instead, described using simulations to better understand the impact of ethical concerns identified elsewhere (Brandao et al., 2020 ; Stormont, 2008 ; Tanzi et al., 2015 ). Only one paper looked at cases of actual robot deployment—although in one of the two cases the setting was highly controlled—to test the introduction of an ethical framework (Amigoni & Schiaffonati, 2018 ).

Fairness and discrimination

Concerns associated with fairness and discrimination were the most frequently discussed ethical considerations in our review (4/6 papers); in three of these papers, discrimination was looked at in terms of disaster victims, in one, instead, as relating to rescue operators. As Amigoni and Schiaffonati point out:

Hazards and benefits should be fairly distributed (…) to avoid the possibility of some subjects incurring only costs while other subjects enjoy only benefits. This condition is particularly critical for search and rescue robot systems, e.g., when a robot makes decisions about prioritizing the order in which the detected victims are reported to the human rescuers or about which detected victim it should try to transport first (Amigoni & Schiaffonati, 2018 ).

In their paper, Brandão and colleagues provide an illuminating practical illustration of this concern (Brandão et al., 2020 ). The authors describe the hypothetical case of a UAV deployed after a disaster to search for victims and deliver medications. After each mission, the UAV needs to recharge its batteries and reload supplies at the base, in the center of the city. Because the distribution of the city’s population, as is often the case, is not uniform in terms of density, age, ethnicity and gender, the authors continue, the UAV’s planned paths will have skewed distribution of these characteristics. So, for instance, if the city in question has a high-density concentration of university students in its center, and the UAV begins its exploration missions from the area surrounding the base station, it will mostly find young people, who are usually more likely to survive than older people living elsewhere. The UAV will therefore be successful in terms of finding as many people as possible, but at the same time it will not respect the notion of distributive fairness, according to which, in this context, priority should be given to those who are most at risk (and need to be found first). Such a robot would confirm or even reinforce common critical views about disaster response missions according to which policies for selecting disaster response locations are often unfair (O’Mathuna et al., 2013 ).

Looking at the impact of information systems used during crisis management and disaster relief, Tanzi and colleagues also emphasize the risk of issues of social justice, pointing out that inclusive design is often lacking in emergency systems and that this may contribute to or exacerbate the marginalization of certain social groups and communities (Tanzi et al., 2015 ).

Carlsen et al., predict that male rescuers, being the ones traditionally involved in the riskiest and most physically demanding rescue operations, may be discriminated as they will be the most likely to be replaced by rescue robots (Carlsen et al., 2015 ).

False or excessive expectations

This theme was only discussed by Harbers and colleagues . , who point out that stakeholders are generally unable to make sound assessments about the capabilities and limitations of rescue robots. In the authors’ view, this inability can lead stakeholders to overestimate or underestimate the capabilities of rescue robots. In the first case, this may translate into unjustified reliance on their performance, and thus, for example, into false hopes that the robots may save certain victims, or into their deployment for tasks for which they are not suitable or under inappropriate conditions. In the second case, when robots’ capabilities are underestimated, they may be underutilized, leading to a waste of precious resources (Harbers et al., 2017 ).

Labor replacement

Both Carlsen et al. and Harbers and colleagues report that stakeholders predict that rescue robots will likely replace human operators in the most physically challenging or high-risk rescue missions. While Carlsen et al. then focus on the likelihood of ensuing discrimination towards male responders, as mentioned above, (Carlsen et al., 2015 ), Harbers and colleagues express a concern that replacing humans with robots may determine degraded performance with respect to victim contact, situation awareness, manipulation capabilities, etc., pointing out that robot-mediated contact with victims may interfere with medical personnel’s ability to perform triage or provide medical advice or support (Harbers et al., 2017 ).

Questions related to privacy are extensively examined in three of the papers we identified. According to Harbers and colleagues, the use of robots generally leads to an increase in information gathering, which can jeopardize the privacy of personal information. This may be personal information about rescue workers, such as images or data about their physical and mental stress levels, but also about victims or people living or working in the disaster area. Harbers et al. add that the loss of privacy potentially associated with the deployment of robots in disaster scenarios does not necessarily result in an ethical dilemma: indeed, given the critical nature of search and rescue operations, the benefits of collecting information in such settings largely outweigh any harms it may cause. This will require, however, that the information gathered by the robots is not shared with anyone outside professional rescue organizations and is exclusively used for rescue purposes. Given the time-critical, data-rich, high-stakes and often quite chaotic conditions that characterize rescue operations, ensuring such careful handling, the authors conclude, will require particular care (Harbers et al., 2017 ).

Reporting previous work (Buescher et al., 2013 ), Tanzi and colleagues emphasize the need for regulation when using information technology in crisis or emergency situations, in order to clarify misunderstandings about situations or cases (i.e. whether it is possible to collect, process and share data with other stakeholders) and to foster good practices. As an illustration, the authors quote Buescher et al., who explained how, during the 2005 London terrorist attack, a failure to share data, legitimacy and silo-thinking led to inefficiencies and mistakes on the part of the emergency agency, due to misinterpretations of the requirements of the UK Data Protection Act of 1998 (Buescher et al., 2013 ; Tanzi et al., 2015 ).

Within navigation planning, Brandão and colleagues explain that ensuring fairness in rescue robot navigation requires collecting data on the distribution of certain features of the population affected by a disaster. This may realistically involve privacy issues with the data collection itself, or with its analysis. It may also lead to leaks coming from data breaches but also from correlations within observed robot behavior, as paths taken by a robot could reveal information about the personal characteristics of the people in the city or other location where the robot is deployed (Brandão et al., 2020 ).

Responsibility

In the paper by Tanzi et al . issues of responsibility are viewed as associated with liability in the event of technical failures or accidents and injuries to victims (Tanzi et al., 2015 ).

Harbers and colleagues instead focus on responsibility assignment problems, which, they say, can apply to both moral and legal responsibility, where moral responsibility concerns blame and legal responsibility, instead, concerns accountability. Such problems, according to the authors, can arise when robots act with no human supervision. If a robot malfunctions, behaves incorrectly, makes a mistake or causes harm, it may be unclear who is responsible for the damage caused: the operator, the software developer, the manufacturer or the robot itself. Responsibility assignment problems, they continue, become particularly complicated when the robot has some degree of autonomy, self-learning capabilities or is capable of making choices that were not explicitly programmed (Harbers et al., 2017 ).

Carlsen and colleagues note that employing robots that are capable of learning introduces a “man in the middle” regarding responsibilities, explaining that, arguably, previous owners, as well as the designers, producers and users of such robots, could be held responsible for any problems they cause (Carlsen et al., 2015 ; Johansson, 2010 ). They also point out that first responders and other operators might be concerned about robots collecting visual data during rescue operations, as this may involve other people being able to watch them closely during missions. This would be a drastic change compared to the current situation, opening up the possibility of rescue operations being evaluated in unprecedented ways (Carlsen et al., 2015 ).

Harbers and colleagues acknowledge that although attention to safety is clearly one of the key priorities than need to be taken into account when deploying rescue robots, this priority will often have to be balanced against other values, as rescue missions necessarily involve safety risks. Certain of these risks can be mitigated by replacing operators with robots, but robots themselves, in turn, may determine other safety risks, mainly because they can malfunction. Even when they perform correctly, robots can still be harmful: they may, for instance, fail to identify and collide into a human being. In addition, robots can hinder the well-being of victims in subtler ways. For example, the authors argue, being trapped under a collapsed building, wounded and lost, and suddenly being confronted with a robot, especially if there are no humans around, can in itself be a shocking experience (Harbers et al., 2017 ).

Focusing specifically on the use of UAVs, Tanzi et al. also emphasize the risks associated with collisions and accidents, pointing out that even high-end military drones like the Predator crash with some frequency, although injuries are rare, and that in urban environments, small UAVs can still cause injury or property damage (Tanzi et al., 2015 ).

Amigoni and Schiaffonati acknowledge that, while risks should be contained as much as possible, it is also evident that a completely risk-free situation is not possible for robot systems operating in search and rescue missions (Amigoni & Schiaffonati, 2018 ).

The question of trust in autonomous systems is the focus of one of the papers identified by our review. In his paper, Stormont highlights how trust by an agent in another agent requires two beliefs: that an agent that can perform a task to help another achieve a goal has a) the ability to perform the task and b) the desire to perform it (Stormont, 2008 ). He then points out that two main components of trust have been identified in the literature: confidence and reputation. Stormont claims that autonomous systems and robots in general tend to not have a good reputation. While Stormont is unable to provide a comprehensive explanation for robots’ reputational problem, he suggests that confidence, the other component of trust, must be involved. In the author’s view, humans lack confidence in autonomous robots because they are unpredictable. Humans working together are generally able to anticipate each other’s actions in a wide range of circumstances—especially if they have trained together, as is the case in rescue crews. Autonomous systems, instead, often surprise even those who designed them, and such unpredictability can be both concerning and unwelcome in dangerous situations like those that are typical of disaster scenarios.

Over the past years the capabilities of rescue robots have vastly improved. As reported by Delmerico and colleagues in their review of the current state and future outlook of rescue robotics, developments in UAVs have led to new applications for aerial platforms, progress in control and actuation now make it possible for legged robots to negotiate tough terrains, and novel human–robot interfaces are improving the ways in which operators can interact with robots (Delmerico et al., 2019 ). Some of these important advances are already used in field-ready commercial products, making widespread adoption of rescue robots increasingly likely. Yet, the ethical concerns associated with the use of such robots remain largely overlooked in the literature, as confirmed by the very limited corpus of research identified by our scoping review. With no relevant publications appearing before 2008 but 5/6 concentrated between 2014 and 2020, there appears to be a modest increase in interest; the annual publication rate, however, remains exceedingly low. Given the abundance of research and guidelines concerned with the ethics of robots, autonomous systems and artificial intelligence that has been produced in recent years (Winfield, 2019 ; Winfield & Jirotka, 2018 ), this finding is somewhat surprising.

Quantitative analysis: anticipating the ethical impact of rescue robots

Most of the papers we identified investigate the ethical concerns associated with rescue robots by envisioning potential scenarios or developing simulations of what could take place, responding to uncertainty with an anticipatory approach. Anticipating the impact of novel technologies is notoriously difficult, as illustrated by the control dilemma formulated by David Collingridge (Collingridge, 1980 ), which remains central in discussions among scholars of technology assessment. The dilemma states:

attempting to control a technology is difficult…because during its early stages, when it can be controlled, not enough can be known about its harmful social consequences to warrant controlling its development; but by the time these consequences are apparent, control has become costly and slow.

Since its formulation, approaches aimed at engaging with the control dilemma have mostly focused on reducing the uncertainty inherent in the early stages of technological development, proposing that technologies should be designed proactively in ways that prevent negative consequences and risks while at the same time striving to achieve positive impacts (van de Poel, 2015 ). This is the approach in a range of methods that have been developed to structure the process of practicing ethics in Research and Innovation, such as Constructive Technology Assessment (Rip et al., 1995 ), Value Sensitive Design (Friedman et al., 2006 ) and Responsible Innovation (Owen et al., 2013 ). Most of the papers identified by our review, indeed, refer to one or another of these methods, displaying an awareness that anticipation and proactive approaches, if not always enough to solve the control dilemma, are arguably a helpful component of the solution. Amigoni and Schiaffonati also discuss the value of coupling anticipation with explorative experiments. Recognizing that technological innovation can defy foresight by behaving in unexpected ways, this approach calls for gradual introduction of novel technologies into society, so that their effects can be monitored, and their design iteratively changed. To answer ethical questions concerning the adoption of rescue robots, Amigoni and Schiaffonati propose, anticipatory methods should be paired with explorative experiments aimed at gaining knowledge on the behavior of such robots in real world deployments (Amigoni & Schiaffonati, 2018 ).

Another finding that emerges in the studies we identified is the incorporation of empirical research, namely explorations of stakeholder views and values, in the ethical assessment of novel technologies. Such empirical explorations, often in the form of qualitative studies, are actually a key element of some of the approaches mentioned earlier. Responsible Innovation, for instance, advocates the involvement of relevant stakeholders throughout the life cycle of research and innovation (Von Schomberg, 2013 ), highlighting the importance of awareness and sensitivity to social and cultural contexts. Likewise, Value Sensitive Design includes empirical investigations using tools from social sciences research to explore the human context in which novel artifacts will function (Friedman et al., 2006 ).

Qualitative analysis: the core ethical themes

The core ethical themes that emerged from the qualitative text analysis of the publications in our review reflect several of the issues that are debated in the wider literature both on ethics in robotics and in disaster ethics. This may explain why, unlike what might be expected, we found limited interplay between these themes and the contexts, robot types and methods used in the studies (see Table 2 ).

Concerns associated with fairness and discrimination are discussed in most of the papers identified and notably by Brandão and colleagues. Many scholars in disaster ethics share the view that individuals most at risk should be given priority (Merin et al., 2010 ; O’Mathuna et al., 2013 ). This view is grounded in the normative position, typical of prioritarian ethical approaches to distributive justice—as Brandão and colleagues note—that those who have a greater need have a stronger moral claim to resources. Thus, in disaster settings, older victims and children should be given priority over victims who are less at risk. Brandão and colleagues’ paper shows that robot navigation planning has implications in terms of distributive justice and indirect discrimination. Indeed, how navigation is planned necessarily modifies the likelihood that certain people can access or not the benefits associated with the presence or actions of the rescue robot itself. This can lead to structural injustices having to do, for instance, with income- or age-related segregations in given urban areas (Brandão et al., 2020 ). Existing work conducted from a Responsible Innovation perspective, Brandão et al. remind us, has highlighted that fairness considerations are crucial to stakeholders when considering the application of autonomous systems (Webb et al., 2018 ). Similarly, questions of fairness, and controversies over lack of fairness (Angwin et al., 2016 ; Chouldechova, 2017 ) have occupied much of the public discourse on AI ethics, and as highlighted by Tanzi and colleagues, the lack of inclusive design can generate issues of social justice (Tanzi et al., 2015 ).

False or excessive expectations and trust

The question of false or excessive expectations is discussed by Harbers and colleagues (Harbers et al., 2017 ), and is closely linked to issues of trust (Stormont, 2008 ).

Popular accounts of the failures and successes of robots, e.g. in the media, in news sources, in science fiction literature and in movies, often mislead public expectations of what robots are and what they can do. In the case of UAVs, for example, before they were widely commercialized, media focus on specific aspects of drone usage generated false impressions and ideas, e.g. that most UAVs were armed Predator-type drones, owned and operated by just a few countries for military purposes (Franke, 2013 ).

Misconceptions about robots may also derive from the fact that robots built to interact with humans often give the impression that they are more intelligent than they really are (Kwon et al., 2016 ). To describe the phenomenon that takes place when humans develop incorrect or unrealistic expectations about the capabilities of complex engineered systems, Kwon and colleagues introduced the term “expectations gap”. They pointed out that robots are built to have specific skills, while humans usually have a wide range of capabilities. Because humans have a tendency to anthropomorphize human-like objects, including robots (Lemaignan et al., 2014 ), they also tend to generalize human mental models to those robots (Dautenhahn, 2002 ) and may overestimate the robot’s actual range of capabilities, at least initially. Human tendencies to misattribute positive human characteristics to robots may result in false expectations and lead to misplaced trust, which can then quickly turn into disappointment and eventually mistrust. If they interfere with teamwork efficiency, false expectations and misplaced trust can have dangerous consequences, particularly when robots support safety-critical tasks, as in the case of search and rescue missions (Groom & Nass, 2007 ). These factors provide a possible explanation for the reputational issues and lack of human confidence that characterize autonomous robots (Stormont, 2008 ).

Carlsen et al. and Harbers and colleagues suggest that search and rescue jobs could eventually be taken over by rescue robots (Carlsen et al., 2015 ; Harbers et al., 2017 ). Both papers report that this concern was raised during workshops with stakeholders, specifically referring to fire-fighters. One current view is that occupations vulnerable to robotization are those that are intensive in routine or predictable tasks, and that fire-fighters, who generally represent a substantial proportion of search and rescue operators, score low on such a vulnerability scale (Owen, 2020 ). Nonetheless, ensuring that rescue robotics is both innovative and beneficial will require a clear understanding of its societal impact, with a goal, among others, of prioritizing innovations that complement rather than replace rescue workers.

The goals and capabilities of rescue robots entail increased information gathering which can determine risks of malicious data exploitation or simply lead to privacy loss for rescue operators, victims or other stakeholders at disaster sites, whose information is purposefully or incidentally collected. As discussed by Harbers and colleagues, this may concern personal information regarding physical and mental stress levels, or images of victims bodies’, or photos of people’s devastated homes (Harbers et al., 2017 ). In addition, as Brandão and colleagues note, data leaks can be generated by security breaches but also through correlations within observed robot behavior (Brandão et al., 2020 ). Therefore, although increased information flow is widely accepted as appropriate for emergencies and disasters (Sanfilippo et al., 2019 ), information flows across robots introduce new complexity and provide more opportunities for privacy infringements. Privacy thus emerges as a key human rights concern in relation to the deployment of rescue robots, requiring careful regulation and good practices.

Coming to the question of responsibility, two interesting concerns emerge from our review: first, responsibility assignment and, second, rescue operators’ potential worry that the presence of robots during rescue missions may increase the transparency of operations to their detriment. Regarding responsibility assignment problems, as Harbers and colleagues point out, the issue may become particularly complicated when the rescue robot has some degree of autonomy, self-learning capabilities or is capable of making choices that were not explicitly programmed (Harbers et al., 2017 ). According to many scholars, humans should always be responsible for what a robot does (Coeckelbergh, 2020 ). Indeed, most global initiatives focusing on ethics in robotics and AI state that autonomous systems should be auditable, to ensure that designers, manufacturers, owners, and operators are held accountable for the technology or system's actions, and are considered responsible for any harm it might cause (Bird et al., 2020 ). In a different view, the possibility of extending legal personality to robots has been proposed as a mechanism that could be used to apply directly to robots obligations that currently apply only to individuals and legal persons such as companies (Fosch-Villaronga, 2019 ). A European Parliament resolution of 2017 thus suggested that the status of “electronic persons” might be conferred to the most sophisticated autonomous robots. The proposal, however, was opposed by scholars who argued that conferring rights and personhood to robots would be “morally unnecessary and legally troublesome”, as difficulties in holding such electronic persons accountable would outweigh any moral interests that such a legal fiction might protect (Bryson et al., 2017 ). Overall, while there is wide agreement that accountability, liability, and the rule of law must be upheld in the face of new technologies (European Group on Ethics in Science and New Technologies, 2018 ), how this should be done and how responsibility should be allocated when it comes to robots with increasing autonomy remains to be defined (Muller, 2020 ). One way forward, it has been suggested, might be to gather more qualitative and quantitative data in order to better understand how likely any harms connected to robot deployments actually might be, and whether such harms would justify the implementation of specific measures (Fosch-Villaronga, 2019 ).

Moving on to the question of what data gathered by robots might reveal about rescue operations and the associated concerns rescuers may have, in many legal systems both professional and volunteer rescuers are shielded from exposure to liability during operations if they discharge their duties reasonably and in good faith. Based on the above, it would be interesting to explore the views of responders who could be involved in missions alongside rescue robots, to better understand any concerns they might have about the information gathered by robots during missions, and to how its sharing might potentially affect them, e.g. in terms of perception and self-perception of their professionalism and role within society.

The account of safety given by the authors of the papers in our review mainly centers on the risk of technical failures or malfunctions leading to collisions and injuries between rescue robots deployed in operative areas and persons on the ground, both responders and victims. Harbers and colleagues however also point out that robots can interfere with victims’ well-being in different ways, noting that the experience of being, for instance, trapped under a collapsed building, wounded and afraid, and suddenly confronted by a robot, may be terrifying (Harbers et al., 2017 ). This suggestion resonates with the concerns laid forth by van Wynsberghe and Comes in their recent paper on ethical considerations about humanitarian drones (van Wynsberghe & Comes, 2020 ), who emphasize the safety concerns having to do with the behavioral, psychological and physiological wellbeing of people experiencing robots.

Conclusions

Taken together, the quantitative and qualitative findings of our scoping review show that, while the ethical concerns in rescue robotics are underexamined in the literature, the papers we identified uniformly endorse a proactive approach to handling such ethical concerns and display an awareness that ethical considerations need to be taken into account before rescue robots become ubiquitous in disaster settings.

As well as providing more in-depth analysis of the issues raised by the publications included here, future research should consider other ethical considerations that might be influential. Findings from van Wynsberghe and Comes's recent study on ethical concerns associated with the use of humanitarian drones, for instance, suggest that dignity, deskilling and informational transparency deserve attention (van Wynsberghe & Comes, 2020 ). In addition, more qualitative work is needed to explore the views of experts and professionals in the search and rescue robotics domain in order to move from hypothetical scenarios and simulations to understandings of lived experience with different types of robots in different contexts. Combining the results of normative and empirical research in the ethics of rescue robotics will clarify the issues that rescuers face when deploying robots in disaster scenarios, while at the same time facilitating the development of practical decision-making tools and empirically-informed guidelines for such deployment (Ienca et al., 2018 ).

The results of our study should be considered in light of certain limitations. First, new records in the academic literature will have been published between the time we concluded our review and publication of this study. Second, as we only selected papers published in English, we may have missed data, analyses and reflections reported in other languages. Finally, we did not include the grey literature, so we cannot rule out that relevant websites, reports, theses, and other documents exist beyond the publications we identified. Despite these limitations, by offering a comprehensive view of the current literature on the ethical concerns associated with the use of robots in disasters, this scoping review provides a helpful starting point for further exploration, analysis and reflection.

Adams, S., Levitan, M., & Friedland, C. (2014). High resolution imagery collection for post-disaster studies utilizing unmanned aircraft systems (UAS). Photogrammetric Engineering & Remote Sensing, 12 , 1161–1168.

Article   Google Scholar  

Amigoni, F., & Schiaffonati, V. (2018). Ethics for robots as experimental technologies: Pairing anticipation with exploration to evaluate the social impact of robotics. IEEE Robotics & Automation Magazine, 25 (1), 30–36.

Angwin, J., Larson, J., Mattu, S., & Kirchner, L. (2016). Machine bias: there’s software used across the country to predict future criminals and it’s biased against blacks . Propublica.

Google Scholar  

Arksey, H., & O’Malley, L. (2005). Scoping studies: Towards a methodological framework. International Journal of Social Research Methodology, 8 (1), 19–32.

Bird, E., Fox-Skelly, J., Jenner, N., Larbey, R., Weitkamp, E., & Winfield, A. (2020). The ethics of artificial intelligence: Issues and initiatives . Brussels. https://www.europarl.europa.eu/RegData/etudes/STUD/2020/634452/EPRS_STU(2020)634452_EN.p%0Adf%3E .

Birk, A., & Carpin, S. (2006). Rescue robotics a crucial milestone on the road to autonomous systems. Advanced Robotics, 20 (5), 595–605.

Brandão, M., Jirotka, M., Webb, H., & Luff, P. (2020). Fair navigation planning: A resource for characterizing and designing fairness in mobile robots. Artificial Intelligence . https://doi.org/10.1016/j.artint.2020.103259

Article   MathSciNet   MATH   Google Scholar  

Bryson, J., Diamantis, M. E., Grant, T. D. (2017). Of, for, and by the people: The legal lacuna of synthetic persons. Artificial Intelligence and Law , 25 , 273–291.

Buescher, M., Wood, L., & Perng, S. (2013). Privacy, security, liberty: informing the design of EMIS. In 10th International ISCRAM Conference , Baden-Baden , Germany , pp. 401–410.

Carlsen, H., Johansson, L., Wikman-Svahn, P., & Dreborg, K. H. (2015). Co-evolutionary scenarios for creative prototyping of future robot systems for civil protection. Technological Forecasting & Social Change, 84 , 93–100.

Chatila, R., Havens, J. C. (2019). The IEEE global initiative on ethics of autonomous and intelligent systems. In: Aldinhas Ferreira, M., J. Silva Sequeira, G. Singh Virk, M. Tokhi & E. Kadar (eds.), Robotics and well-being. Intelligent systems, control and automation: Science and engineering , vol 95. Cham: Springer. https://doi.org/10.1007/978-3-030-12524-0_2

Chouldechova, A. (2017). Fair prediction with disparate impact: A study of bias in recidivism prediction instruments. Big Data, 5 (2), 153–163.

Coeckelbergh, M. (2020). AI ethics . MIT Press, Cambridge, MA. ISBN: 9780262538190.

Collingridge, D. (1980). The social control of technology . Pinter.

Colquhoun, H., Levac, D., & O’Brien, K. (2014). Scoping reviews: Time for clarity in definition, methods and reporting. Journal of Clinical Epidemiology, 67 , 1291–1294. https://doi.org/10.1080/1364557032000119616

Corbin, J., & Strauss, A. (2008). Basics of qualitative research: Techniques and procedures for developing grounded theory (3rd edn). Sage Publications, Inc. https://doi.org/10.4135/9781452230153 .

Cuny, F. (1992). Introduction to disaster management lesson 1: The scope of disaster management. Prehospital and Disaster Medicine, 7 (4), 400–409.

Dautenhahn, K. (2002). Design spaces and niche spaces of believable social robots.In Proceedings of the 11th IEEE International Workshop on Robot and Human Interactive Communication (pp. 192–197). IEEE.

Delmerico, J., Mintchev, S., Giusti, A., Gromov, B., Melo, K., Horvat, T., Cadena, C., Hutter, M., Ijspeert, A., Floreano, D., Gambardella, L. M., Siegwart, R., & Scaramuzza, D. (2019). The current state and future outlook of rescue robotics. Journal of Field Robotics, 36 , 1171–1191.

Devault, J. (2000). Robotic system for underwater inspection of bridge piers. IEEE Instrumentation & Measurement Magazine, 3 , 32–37.

Ellenberg, A., Branco, L., Krick, A., Bartoli, I., & Kontsos, A. (2015). Use of unmanned aerial vehicle for quantitative infrastructure evaluation. Journal of Infrastructure Systems, 21 , 04014054.

European Group on Ethics in Science and New Technologies. (2018). Statement on artificial intelligence, robotics and autonomous systems. European Commission, Directorate-General for Research and Innovation, Unit RTD.01. https://op.europa.eu/en/publication-detail/-/publication/dfebe62e-4ce9-11e8-be1d-01aa75ed71a1 .

Fosch-Villaronga, E. (2019). Robots, healthcare, and the law: Regulating automation in personal care (1st ed.). Routledge.

Franke, U.E. (2013). The five most common media misrepresentations of UAVs. In Aaronson, M., A. Johnson (eds.), Hitting the target? how new capabilities are shaping international intervention (pp. 19–31). London: Royal United Services Institute.

Friedman, B., Kahn, P., & Borning, A. (2006). Value sensitive design and information systems. In P. Zhang & D. Galletta (Eds.), Human-computer interaction in management information systems: Foundations (pp. 348–372). Sharpe.

Grant, M., & Booth, A. (2009). A typology of reviews: An analysis of 14 review types and associated methodologies. Health Information & Libraries Journal, 26 (2), 91–108.

Green, B. L., Friedman, M. J., de Jong, J., Keane, T. M., Fairbank, J. A., Solomon, S. D., Donelan, B., & Frey-Wouters, E. (Eds.). (2003). Trauma interventions in war and peace: Prevention, practice, and policy . Kluwer Academic Publishers.

Groom, V., & Nass, C. (2007). Can robots be teammates? Benchmarks in human-robot teams. Interaction Studies, 8 (3), 483–500.

Gustavsson, M. E., Arnberg, F. K., Juth, N., & von Schreeb, J. (2020). Moral distress among disaster responders: What is it? Prehospital and Disaster Medicine, 35 (2), 212–219. https://doi.org/10.1017/S1049023X20000096

Harbers, M., de Greeff, J., Kruijff-Korbayová, I., Neerincx, M. A., & Hindriks, K. V. (2017). Exploring the ethical landscape of robot-assisted search and rescue . Cham: Springer. https://doi.org/10.1007/978-3-319-46667-5_7

Book   Google Scholar  

Ienca, M., Ferretti, A., Hurst, S., Puhan, M., Lovis, C., Vayena, E. (2018). Considerations for ethics review of big data health research: A scoping review. PLOS One , 13 (10).

Johansson, L. (2010). The functional morality of robots. International Journal Technoethics , 1 , 65-73.

Kochersberger, K., Kroeger, K., Krawiec, B., Brewer, E., & Weber, T. (2014). Post-disaster remote sensing and sampling via an autonomous helicopter. Journal of Field Robotics, 31 , 510–521.

Kwon, M., Jung, M., & Knepper, R. (2016). Human expectations of social robots. In HRI ’16: The Eleventh ACM/IEEE International Conference on Human Robot Interaction (pp. 463–464). IEEE.

Lattanzi, D., & Miller, G. (2017). Review of robotic infrastructure inspection systems. Journal of Infrastructure Systems, 23 , 04017004.

Lemaignan, S., Fink, J., & Dillenbourg, P. (2014). The dynamics of anthropomorphism in robotics. In Proceedings of the 2014 ACM/IEEE International Conference on Human-Robot Interaction (pp. 226–227). ACM.

Levac, D., Colquhoun, H., & O’Brien, K. (2010). Scoping studies: Advancing the methodology. Implementation Science, 5 (1), 69.

Lichoki, P., Billard, A., & Kahn, P., Jr. (2011). The ethical landscape of robotics. Robotics & Automation Magazine, IEEE, 18 (1), 39–50.

Merin, O., Ash, N., Levy, G., Schwaber, M., & Kreiss, Y. (2010). The Israeli field hostpital in Haiti—ethical dilemmas in early disaster response. New England Journal of Medicine, 362 (11), e38.

Muller, V. (2020). Ethics of artificial intelligence and robotics. In The stanford encyclopedia of philosophy . https://plato.stanford.edu/archives/win2020/entries/ethics-ai/ .

Murphy, R. R. (2014). Disaster robotics . MIT Press.

Murphy, R., Kravitz, J., Stover, S., & Shoureshi, R. (2009). Mobile robotics in mine rescue and recovery. IEEE Robotics and Automation Magazine, 16 , 91–103.

Murphy, R., & Stover, S. (2007). Rescue robots for mudslides: A descriptive study of the 2005 La Conchita mudslide response. Journal of Field Robotics, 25 , 3–16.

O’Mathuna, D., Gordijn, B., & Clarke, M. (2013). Disaster bioethics: Normative issues when nothing is normal . Springer Science and Business Media.

Owen, E. (2020). Firms vs. workers? The political economy of labor in an Era of global production and automation. Working Paper. https://www.internationalpoliticaleconomysociety.org/sites/default/files/paper-uploads/[email protected] .

Owen, R., Bessant, J., & Heintz, M. (2013). Responsible innovation: Managing the responsible emergence of science and innovation in society . Wiley.

Pham, M. T., Rajic, A., Greig, J. D., Sargeant, J. M., Papadopoulos, A., & McEwen, S. A. (2014). A scoping review of scoping reviews: Advancing the approach and enhancing the consistency. Research Synthesis Methods, 5 , 371–385. https://doi.org/10.1002/jrsm.1123

Recchiuto, C. T., & Sgorbissa, A. (2017). Post-disaster assessment with unmanned aerial vehicles: A survey on practical implementations and research approaches. Journal of Field Robotics . https://doi.org/10.1002/rob.21756a .

Rip, A., Misa, T., & Schot, J. (Eds.). (1995). Managing technology in society. The approach of constructive technology assessment . Pinter.

Rumrill, P., Fitzgerald, S., & Merchant, W. (2010). Using scoping literature reviews as a means of understanding and interpreting existing literature. Work, 35 , 399–404.

Sanfilippo, M., Shvartzshnaider, Y., Reyes, I., Nissenbaum, H., & Egelman, S. (2019). Disaster privacy/privacy disaster. The 47th Research Conference on Communication, Information and Internet Policy. SSRN Electronic Journal. https://doi.org/10.2139/ssrn.3427562

Sharkey, A., & Sharkey, N. (2012). Granny and the robots: Ethical issues in robot care for the elderly. Ethics and Information Technology, 14 (1), 27–40. https://doi.org/10.1007/s10676-010-9234-6

Sharkey, N., & Sharkey, A. (2011). The rights and wrongs of robot care. Robot ethics: The ethical and social implications of robotics . Springer.

Sparrow, R. (2016). Robots in aged care: A dystopian future? AI & Society, 31 (4), 445–454.

Stefanov, W., & Evans, C. (2014). The international space station: a unique platform for remote sensing of natural disasters. Technical Report 20150003831, NASA.

Steimle, E., Murphy, R., Lindemuth, M., & Hall, M. (2009). Unmanned marine vehicle use at Hurricanes Wilma and Ike. Unmanned marine vehicle use at Hurricanes Wilma and Ike. OCEANS (pp. 1–6). IEEE.

Stormont, D. P. (2008). Analyzing human trust of autonomous systems in hazardous environments. AAAI Workshop - Technical Report , WS - 08-05 , 27–32.

Tanzi, T., Sebastien, O., & Rizza, C. (2015). Designing autonomous crawling equipment to detect personal connected devices and support rescue operations: Technical and societal concerns. Radio Science Bulletin . https://doi.org/10.23919/URSIRSB.2015.7909472

International Council of Nurses. (2012). The ICN Code of ethics for nurses . http://www.icn.ch/who-we-are/code-of-ethics-for-nurses/ .

Torok, M., Golparvar-Fard, M., & Kochersberger, K. (2014). Image-based automated 3D crack detection for post-disaster building assessment. Journal of Computing in Civil Engineering, 28 , A4014004.

van de Poel, I. (2015). An ethical framework for evaluating experimental technology. Science and Engineering Ethics, 22 , 667–686.

van Wynsberghe, A., & Comes, T. (2020). Drones in humanitarian contexts, robot ethics, and the human–robot interaction. Ethics in Information Technology, 22 , 43–45. https://doi.org/10.1007/s10676-019-09514-1

Von Schomberg, R. (2013). A vision of responsible research and innovation. Responsible Innovation: Managing the Responsible Emergence of Science and Innovation in Society, 26 , 51–74.

Webb, H., Koene, A., Patel, M., & Vallejos, E. (2018). Multi-stakeholder dialogue for policy recommendations on algorithmic fairness.In Proceedings of the 9th International Conference on Social Media and Society (pp. 395–399).

Williams, J. R. (2015). Medical ethics manual , 3rd edition. World Medical Association.

Winfield, A. (2019). An updated round up of ethical principles of robotics and AI. Nature Electronics, 2 , 46–48.

Winfield, A. N. T., & Jirotka, M. (2018). Ethical governance is essential to building trust in robotics and artificial intelligence systems. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences . https://doi.org/10.1098/rsta.2018.0085

Zuzanek, P., Zimmermann, K., & Hlavac, V. (2014). Accepted autonomy for search and rescue robotics. First International Workshop, MESAS 2014, Rome, Italy, May 5-6 2014 , 231. Springer.

Download references

Open access funding provided by Università degli Studi di Genova within the CRUI-CARE Agreement. This work was supported by the Italian Ministry of Education, University and Research through the DIONISO project (Research code: B82C12000070001; National Civil Service Code SCN_00320), under the Smart Cities and Communities and Social Innovation program.

Author information

Authors and affiliations.

Department of Informatics, Bioengineering, Robotics and Systems Engineering, University of Genoa, Genoa, Italy

Linda Battistuzzi, Carmine Tommaso Recchiuto & Antonio Sgorbissa

You can also search for this author in PubMed   Google Scholar

Corresponding author

Correspondence to Linda Battistuzzi .

Ethics declarations

Conflict of interest.

The authors declare that they have no conflicts of interest.

Additional information

Publisher's note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ .

Reprints and permissions

About this article

Battistuzzi, L., Recchiuto, C.T. & Sgorbissa, A. Ethical concerns in rescue robotics: a scoping review. Ethics Inf Technol 23 , 863–875 (2021). https://doi.org/10.1007/s10676-021-09603-0

Download citation

Accepted : 21 June 2021

Published : 02 July 2021

Issue Date : December 2021

DOI : https://doi.org/10.1007/s10676-021-09603-0

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Robot ethics
• Rescue robotics
• Disaster robotics
• Search and rescue
• Scoping review
• Find a journal
• Publish with us
• Track your research

• Research article
• Open access
• Published: 05 December 2018

Search and rescue with autonomous flying robots through behavior-based cooperative intelligence

• Ross D. Arnold   ORCID: orcid.org/0000-0003-1915-5857 1 ,
• Hiroyuki Yamaguchi 2 &
• Toshiyuki Tanaka 2  

Journal of International Humanitarian Action volume  3 , Article number:  18 ( 2018 ) Cite this article

23k Accesses

97 Citations

71 Altmetric

Metrics details

A Correction to this article was published on 01 June 2019

This article has been updated

A swarm of autonomous flying robots is implemented in simulation to cooperatively gather situational awareness data during the first few hours after a major natural disaster. In computer simulations, the swarm is successful in locating over 90% of survivors in less than an hour. The swarm is controlled by new sets of reactive behaviors which are presented and evaluated. The reactive behaviors integrate collision avoidance, battery recharge, formation control, altitude maintenance, and a variety of search methods to optimize the coverage area of camera and heart-beat locator sensors mounted on the robots. The behaviors are implemented in simulation on swarms of sizes from 1 to 20 robots. The simulation uses actual location data, including post-disaster satellite imagery, real locations of damaged and inundated buildings, and realistic victim locations based on personal interviews and accounts. The results demonstrate the value of using behavior-based swarming algorithms to control autonomous unmanned aerial vehicles for post-disaster search and assessment. Three examples of algorithms that have been effective in simulation are presented .

Introduction

WITH little warning, a powerful earthquake shatters the quiet calm of a coastal city, followed shortly by the periodic waves of a brutal tsunami strike. Within minutes, local rescue workers rush to disaster sites, where they are greeted with a morass of broken buildings, piled cars, and splintered debris. Where once streets and fields stretched peacefully, now sit water-inundated lagoons filled with hazardous material. Mobility is extremely limited. Conditions are harsh; it is cold, night is soon to fall, and it is starting to snow. There are debris everywhere; it is hard to even walk.

The workers pull their truck up to a roadblock of over-turned cars. Only a half dozen workers have made it to the site so far. But people are in the water, trapped in cars, trapped in buildings, and there is no time to wait. The rescue workers pull small, cheap quadcopter unmanned aerial vehicles (UAVs) out from the back of their truck. The workers are already cold and wet, thinking about finding casualties, and preparing equipment. They just want to know where to find people, but how can they find anyone in this devastation?

Fumbling with the UAVs, wearing fireman’s gloves, they manage to start flipping the UAVs on. They pause for a moment, trying to remember how to make the things work. But, they do not have to remember. As soon as they are turned on, the UAVs immediately launch and begin their search automatically. Remembering the apps on their mobile phones, the workers open up their “UAV Search” applications. Immediately, an overhead picture of the scenario appears on a map on their phones – it’s the camera feed from the first UAV.

While two of the workers are looking at their phones, a third and fourth are flipping on more UAVs. Three of the UAVs do not even turn on. They must have been damaged somehow. But it does not matter, seven were able to launch. One by one the UAVs fly up into the sky, flock together, and begin a systematic, targeted search of the inundated regions. At first, workers can only see the camera feeds from each of the UAVs. Able to see several feeds on their screen at once, the workers start look up to see where people are. Motion catches their eyes - there, in the top of the parking garage – a group of 12, waving their hands. The workers radio in for a helicopter, targeting the garage.

Then, among the swarm of cheaper UAVs, a better-equipped one is launched. Then another. Suddenly, on the screen, red dots appear. From the “UAV Search” app, a list of locations appears on the left side, organized from highest to lowest probability of a find, by number of people. As the UAVs continue their search, more and more locations are added. The UAVs move in and out of formation as they locate survivors. One worker clicks on the top find. A snapshot of the camera feed at the time of the find is displayed, along with an arrow pointing from the launch site to the location, and a distance measurement. Immediately, the workers know which direction to go, how far to go, and what the site looks like from the air. Seeing that the location is fallen building with no visible signs of a survivor, two rescue workers immediately set out in that direction, knowing the survivor is likely buried in the rubble.

The vignette above is a fictional “what-if” scenario based on real accounts of the 2011 Great Eastern Japan Earthquake and Tsunami (Editorial Office of the Ishinomaki Kahoku 2014 ). The purpose of the vignette is to share a vision of what could be a significant improvement to post-disaster search and rescue efforts by leveraging teams of autonomous flying robots.

Many sources indicate that the first 72 h of a rescue operation is the most critical (Erdelj et al. 2017 ) (Tait Communications 2012 ), though some studies reduce this window to 48 or even 24 h (Bartels et al. n.d. ). According to analyses of the 2011 Tōhoku tsunami in Japan, the first 24 h was the most critical (Editorial Office of the Ishinomaki Kahoku 2014 ). Studies across more than 1000 SAR missions show a survival rate dropping exponentially during the first 18 h after the onset of SAR efforts, dropping to a survival rate that levels off near 0% after 20 h (Adams et al. 2007 ).

Despite data showing that a concentrated effort to rescue trapped persons during the first few hours after a disaster would likely yield greater effect than any effort made later (Alley 1992 ) (Macintyre et al. 2006 ), these efforts are significantly hampered by lack of situational awareness (Editorial Office of the Ishinomaki Kahoku 2014 ) (Ochoa and Santos 2015 ) (Shimanski 2005 ). Indeed, the lack of situational awareness within this critical time frame is one of the most significant problems immediately following a natural disaster (Ochoa and Santos 2015 ) (Shimanski 2005 ) (Riley and Endsley 2004 ). Aid workers cannot rescue survivors if they do not know where survivors are.

Situational awareness, in this context, is the degree to which aid workers are aware of the state of the disaster environment. This state may include locations of survivors, wreckage, roads, weather, water and other hazards, or any other environmental factor that might affect the rescue effort. Situational awareness has been studied and applied in many different military, civil, commercial, and aerospace applications over the past several decades. Emergency services focus on situational awareness as a key factor in reducing risk and increasing safety, especially in disaster search and rescue situations (Shimanski 2005 ).

Rescue efforts are further hindered by lack of a trained, standing force of aid workers capable of handling the often-huge workload after a major disaster (Alley 1992 ). This is a challenging problem, as the logistical difficulties inherent to maintaining a highly trained standing workforce capable of handling mass-casualty natural disasters are numerous. The approach described in this article directly addresses these issues and, in particular, the situational awareness problem within the critical 20–24-h time frame using an automated, technical solution.

This article presents an approach to disaster search and rescue, data acquisition, and other types of post-disaster assessment using one or multiple heterogeneous autonomous UAVs. The robots work cooperatively as a swarm while controlled by behavior-based artificial intelligence (also called reactive AI). This research combines behavior-based artificial intelligence, swarm intelligence, pattern search theory, and existing disaster data into a theory of improved search and rescue through the use of autonomous flying robots, also called drones, Unmanned Aerial Vehicles (UAV), or Unmanned Aerial Systems (UAS).

Simulation results generated during the research show the approach described in this article to be both effective and time-efficient. The data show that a swarm of just five UAVs with standard parameters Footnote 1 equipped with the software and algorithms developed in this research can consistently achieve a 90% standard sensor coverage rate Footnote 2 over a 2 km 2 area in under 90 min, reaching nearly 99% coverage rate in under 2 h when operating in environments modeled after real tsunami disaster locations. The research shows that it is possible to search a wide range of area in a short time using a swarm of low-cost UAVs. The area can be searched continuously even if one or multiple UAVs in the swarm fail or crash. The swarm requires minimal operator input, freeing up rescue workers for other tasks. Performance using this method, measured as sensor coverage at a certain range over time, is improved compared to existing methods. Ultimately, this approach allows more data to be acquired faster, with less effort, than existing methods.

Actual data regarding the time it takes rescue workers to thoroughly search an area of 2 km 2 after a disaster without the use of UAVs varies greatly and is difficult to quantify. Moreover, it is impossible to say how many non-surviving victims may have survived, had they been found sooner. However, interviews suggest it can take days to search the most significantly affected areas (Editorial Office of the Ishinomaki Kahoku 2014 ). Although the use of individual, separately controlled UAVs is certainly an improvement over no use of UAVs, separately controlled UAVs require constant operator involvement and can still take many hours to achieve a high level of sensor coverage. Therefore, although direct quantitative comparison to existing methods is difficult to make, qualitative assessment supports the conclusion that the approach described in this article is likely to improve access to post-disaster assessment data by a significant margin over existing methods. Whether existing methods take 6 h, 12 h, or 3 days to cover 90% of the disaster area, the 1.5-h benchmark achieved by the five-UAV swarm in our simulation is significantly faster than any of these measures.

The emergence of complex traits and behaviors from interconnected sets of individual parts is a well-researched and documented phenomenon (Arnold and Wade 2015 ) (Koffka 1922 ) (Wiener 1948 ). The use of this phenomenon to create decentralized artificial intelligence (AI) in the control of robots was thoroughly described by Brooks (Brooks 1999 ). Brooks approaches artificial intelligence from the “bottom-up” by investigating the emergent intelligent patterns of robots equipped with individual, simple behaviors. These robots do not possess centralized control; rather, they react to stimuli (in the form of sensor input) in a variety of relatively simple ways. From these simple interactions, intelligent behavior emerges. This approach is known as behavior-based artificial intelligence. In behavior-based AI, a robot’s intelligence is based on a set of relatively simple, independent behaviors, rather than on a centralized control unit.

Brooks implements behavior-based artificial intelligence theory using an architecture he calls the “subsumption architecture.” In his work, robots’ behaviors “subsume” each other depending on the results of a variety of inputs, such as sonar and pressure sensor data. Only one behavior will be active at any given time. The active behavior varies based on sensor data. Brooks successfully implemented this architecture on a variety of applications requiring artificial intelligence, such as navigation and motor control (Brooks 1999 ). The subsumption architecture can be considered one implementation of behavior-based artificial intelligence, which is itself a broader concept.

The behavior-based approach was applied to research on swarm intelligence by Kennedy and Eberhart (Kennedy et al. 2001 ). Swarm intelligence is the resultant intelligent behavior of groups of independent heterogeneous entities behaving as a single system, such as a flock of birds, swarm of ants, or a hive of bees. Individually, the entities in the swarm may not have an understanding of the workings of the system as a whole. There may not be a single focal point of control over the swarm. However, in some way, the swarm still manages to work together as a single system to accomplish a goal. An ant swarm finds food sources, gathers food, and even builds complex structures at times. A flock of birds avoids predators and successfully migrates. Bees gather nectar for the hive over a wide range of conditions and environments. Theories of behavior-based, or reactive, intelligence apply to these swarms of entities. Swarms often function in an intelligent manner through the reactive behaviors implemented by their entities. Through the reactive behaviors of many individual entities, intelligence emerges (Kennedy et al. 2001 ).

Behavior-based formation control was applied to groups of robots by Balch and Arkin (Balch and Arkin 1998 ). They successfully integrated formation behaviors with navigation and hazard avoidance both in simulation and on a set of land-based ground vehicles. The robots’ speeds and turn directions were influenced through a system of votes based on sensory inputs and communication between robots in the group. Several other related papers on formation control for groups of robots were published around the same time frame (Balch and Arkin 1998 ).

Virágh and Vásárhelyi applied principles of flocking behavior to UAVs (Virágh et al. 2014 ) (Vásárhelyi et al. 2014 ). Virágh applied agent-based models to the control of flocks of UAVs, incorporating principles of time delay in communication as well as inaccuracy of onboard sensors. Two decentralized algorithms are proposed in their research: one based on the collective motion of a flock, the other based on collective target tracking. A principle of their research is to use a realistic simulation framework to study the group behavior of autonomous robots.

Swarm algorithms for controlling groups of UAVs are also under exploration for defense systems by the US Department of Defense (Frelinger et al. 1998 ). Their purposes range from combat search and rescue to ballistic missile defeat, in which many of the fundamental techniques used for targeting in defense systems are similar in principle to disaster search and rescue. In both scenarios, swarms of UAVs build upon cooperative behavior-based intelligence to efficiently locate one more multiple targets.

A team from the Naval Postgraduate School designed a swarm control framework called the Service Academy Swarm Challenge (SASC) architecture. The SASC architecture is used to control swarms of heterogeneous robots using the C++ and Python programming languages. SASC has undergone successful field tests deployed on swarms of fixed-wing and quadrotor UAVs.

Additionally, a programming language called Buzz has been specifically designed to facilitate heterogeneous swarm robotics (Pinciroli and Beltrame  2016 ). Buzz allows behaviors to be defined from the perspective of a single robot or from the overall swarm. This programming language is capable of running on top of other frameworks and can be extended to add new types of robots.

For the purpose of disaster search and rescue, behavior-based control of land-based robots was implemented in the HELIOS system (Guarnieri et al. 2009 ). The HELIOS system consists of five land-based, tracked robots used for urban search and rescue. Two of the robots are equipped with manipulators to perform physical tasks, and the other three are equipped with cameras and laser range finders and are utilized to create virtual maps of the environment. The robots can be used separately or as a team for more complex missions. The three robots equipped with laser range finders can move autonomously in unknown environments using a collaborative positioning system. The system as a whole requires control by a human operator.

The use of unmanned aerial systems in search and rescue is an area of high interest (Erdelj et al. 2017 ) (Molina et al. 2012 ) under consideration by a number of high profile organizations, including the American Red Cross, NASA, and the Japanese Ministry of Defense (American Red Cross 2015 ). Many efforts in this area have included the use of individually piloted UAVs, rather than autonomous swarms of robots (Erdelj et al. 2017 ). For example, the European CLOSE-SEARCH project includes the deployment of a single UAV with a ground-based control station to locate someone lost outdoors (Molina et al. 2012 ). The value of UAVs for information-gathering and situational awareness acquisition has been expressed by a number of sources (Erdelj et al. 2017 ) (Molina et al. 2012 ) (American Red Cross 2015 ). Researchers at Carnegie Mellon are investigating the use of swarms of tiny UAVs to map the interiors of buildings after disasters (Williams 2015 ). However, research into the use of swarms of autonomous UAVs to aid in locating survivors during exterior search and rescue appears to be minimal.

Although UAVs and Unmanned Ground Vehicles (UGVs) are already in use for disaster search and rescue (Erdelj et al. 2017 ) (Molina et al. 2012 ) (American Red Cross 2015 ), the use of swarms of UAVs optimized to autonomously cover a disaster area, streaming useful data to operators and each other while avoiding collisions, weaving over and around obstacles, and returning to charge batteries, has been largely absent. This absence seems to be due to a combination of air traffic regulations, laws restricting the use of UAVs, and technical limitations which, until recently, have been difficult to overcome.

Due to these challenges, the control of autonomous swarms of UAVs is a relatively new phenomenon. The Naval Post-Graduate School in Monterey, California, flew a swarm of 50 UAVs controlled by a single operator in 2015 as part of their Zephyr system. At the time, this event is believed to have set the world record for the most UAVs under single operator control (Hambling, 2015 ). The use of swarms of UAVs to aid in post-disaster assessment was imagined in 2016, in a report describing a human-machine interface to control the UAV swarm.

The Orchid disaster response system under development by the UK appears to be the closest to the approach described in this article (Ramchurn et al. 2016 ). It uses decentralized control of a swarm of UAVs to enhance disaster rescue efforts. The Orchid system is designed to interpret crowd-sourced data, building a picture of a situation and providing recommendations for resource allocation. In contrast, this article describes behavior sets and algorithms used to control UAVs to maximize sensor coverage over areas of land and water. This article also presents the results of simulated time trials using swarms of UAVs. The UAVs are controlled by three different behavior sets to search a realistically designed post-disaster location. Data of this particular nature does not appear to be present in the literature.

Distributed coordination is key to enhancing the scope and level of detail of post-disaster assessment. By distributing the workload among many units, the amount of work and the time it takes to do the work is significantly reduced. This also allows scaling the system to larger or smaller areas by simply adding or subtracting units from the swarm. Controlling these individual units through behavior-based artificial intelligence allows them to react successfully to a variety of challenging, changing situations with minimal or no operator input. The behavior-based method of robot control has been a staple of robotics for the last several decades and has a proven track record of success.

Recent technological developments have made modern UAVs more capable and cost-effective, enabling the use of coordinated swarms at reasonable cost. Footnote 3 UAVs can be equipped with built-in hover and maneuver capabilities as well as high definition (HD) and/or infrared (IR) cameras, wireless capabilities to stream live data, and the ability to carry small payloads or additional sensors. This combination of traits has now enabled the practical use of swarms of small, cost-effective UAVs for post-disaster assessment. In order to propel these efforts forward, it is important to demonstrate the significant time-saving effects that the use of such swarms can produce in post-disaster situations. Furthermore, developing and assessing different algorithms to control the swarm as a single, distributed system while also maintaining the individual capability of each separate unit in the swarm is key to the success of this type of system on the whole.

The research described in this article applies the concepts of behavior-based and swarm-based intelligence to control groups of UAVs to locate survivors in disaster search and rescue scenarios. By using data gathered from town records, in-person interviews, survey data, and site visits, several scenarios were built out that depict the post-tsunami environment in 2011 Sendai City, Japan, with a large degree of accuracy. The heights and placement of structures are accurate, and the locations and behaviors of survivors within the scenario are based on real accounts (Editorial Office of the Ishinomaki Kahoku 2014 ) (Municipal Development Policy Bureau 2017 ) (Post-Disaster Reconstruction Bureau 2015 ) (Sato 2015 ) (The Center for Remembering 3.11 2015 ) (Tohoku Regional Development Association n.d. ).

The algorithms used in this research allow the UAVs to dynamically respond to changes in the environment, as well as unknown scenarios and unforeseen circumstances. For example, sensors can malfunction and the UAVs will still retain some measure of utility. A building can be “dropped” in front of a UAV in the simulation, and the UAV will successfully navigate around or over the building, then continue its task.

A dynamically changing environment is a key part of a disaster scenario. Unless injured or safe, survivors do not often stay still. People move to higher floors in buildings. They move towards lights, sounds, higher ground, helicopters, and safety (Editorial Office of the Ishinomaki Kahoku 2014 ). The weather gets cold, it may start to snow or rain, and the sun may go down (Editorial Office of the Ishinomaki Kahoku 2014 ). Night falls, day breaks, visibility changes. Any rescue approach needs to have the flexibility to accommodate these dynamic changes and respond to unknown environments. Our approach demonstrates this flexibility.

A swarm of standard, commercially available autonomous UAVs controlled by behavior-based, cooperative artificial intelligence software may significantly improve the data set containing known victim locations during disaster search and rescue efforts with minimal operator input required. For the purposes of this research, several requirements are imposed on the algorithm sets used to achieve this hypothesis. The intent of these requirements is to provide a practical, flexible system:

Performance —Gather more data faster

Achieve a simulated standard sensor coverage (30 m range) of 90% across 2 km 2 within 24 h.

Achieve a simulated precise sensor coverage (15 m range) of 90% across 2 km 2 within 24 h using a simulated, miniaturized FINDER sensor. Footnote 4

Scalability —Support any number of robots

Supports an arbitrary number of UAVs in the swarm. Due to computational limits during simulation executions, a maximum of 20 of UAVs was used in this research.

Heterogeneity —Support mixed groups of robots and sensor configurations

Different capabilities and sensor configurations supported within the same swarm.

Different UAV types and models supported within the same swarm.

Behavior-based artificial intelligence

Behavior-based artificial intelligence is the concept that intelligence can emerge through the interactions of simple, individual behaviors lacking centralized control. Combining several well-defined but separate behaviors can result in the emergence of intelligent systemic behavior. When used in software and robotics, this approach can provide a high level of robustness, as failed behaviors can be ignored while default behaviors are activated (Brooks 1999 ). The division of logic between behavior modules can allow the system to scale to a high level of complexity without imposing an unmanageable cognitive load on software developers.

Although there are many ways to design robust systems, systems designed with a behavior-based approach to AI are well-suited to reacting to environments dynamically based on sensor inputs without prior knowledge (Brooks 1999 ). These properties are highly desirable in a post-disaster assessment system operating in a volatile environment where the failure of individual parts of a system may be common due to hazardous external factors.

Proposed technique

To enhance post-disaster assessment, search and rescue, and information gathering, we propose using a technique that combines behavior-based artificial intelligence with cooperative swarm behavior. Individual units of a swarm equipped with behavior-based AI are well-suited to perform cooperative tasks (Kennedy et al. 2001 ), as the results of their own behaviors combine together to emerge as individual unit behaviors, and these unit behaviors combine together to emerge as collective swarm intelligence (systemic behavior).

We implement behavior-based AI and cooperative behavior in a simulated swarm of UAVs to search for disaster survivors in a post-disaster environment. We measure the effectiveness of the approach by recording the detection rates over time of the survivors by the swarm. Our goal is to reach a 90% detection rate in under 24 h in simulation.

This approach can be applied to any sort of information gathering and is not limited to just search and rescue. However, using search and rescue gives a direct, tangible way to understand the benefits and effectiveness of the approach.

Proposed algorithms and control methods

To enhance survivor detection through the use of UAV swarms, several control methods are considered (Fig.  1 ). These methods are all implementations of behavior-based AI. Each control method, also referred to as a method or an algorithm , is simply a set of ordered behaviors conceived of and developed during the research. The order of the behaviors within each method is critical as it determines the priority level at which they are executed. As behaviors can be grouped and ordered in many different ways, it is important to figure out which set of behaviors, and in which order, is most effective. The three sets of behaviors (methods) were selected based on the anticipated effectiveness of each set of behaviors as determined by the researchers.

Standard method —UAVs all follow the same pattern.

Spiral method —Upon locating a “critical mass” concentration of survivors, a single UAV moves outward in a spiral pattern, then returns to previous search method.

Scatter method —Each UAV simultaneously moves to a different location in the search pattern.

UAV search methods. Actual patterns are more complex; the patterns depicted here are simplified for clarity. Blue dots are UAVs, gray areas are destination targets, and red triangle is a concentration of survivors. From left to right: standard, spiral, and scatter

The behaviors in the behavior-based software architecture used in this research are all original and were conceived of and created by the researchers. They are implemented as separate, named, plug-and-play software modules. Each of the three control methods consists of some subset of the following 12 behavior modules. These modules are described in detail in the “ Method implementation ” section and briefly here:

Launch —Take off from a stationary position

Avoid —Avoid collisions with buildings and obstacles

Climb —Climb over obstacles

Recharge —Recharge batteries

Height —Maintain a certain height above the ground, buildings, or large objects

Spiral —Move out in an expanding spiral

Form —Maintain distance between other UAVs

Repel —Move away from other UAVs when too close

Seek —Move directly to a specified GPS location

Waypoint —Move towards a preset pattern of waypoints

Scatter —Move individually towards an unallocated waypoint among a set

Wander —Choose a random location and move towards it

These behaviors were conceived based on deductive reasoning, literature search (Brooks 1999 ) (Kennedy et al. 2001 ), and extensive trial and error in simulation. Each behavior is assigned a priority. The UAV control software arranges priorities by the order the behavior modules are loaded into the software. Earlier behaviors, when triggered, prevent later behaviors from occurring at the same time. That is to say, if the avoid behavior is active at a given time, no behaviors at a lower priority than avoid in the list will be activated (such as height or recharge ). A given time in this situation refers to a given tick in the software, which is approximately 15–16 ms. This measure is consistent with the duration of a tick used in personal computers running Microsoft Windows, Apple macOS, or Linux, and mobile operating systems used in UAVs such as the Google Android operating system and iOS.

The UAV re-checks its sensor input at a rate of roughly 60 Hz (or 60 frames per second), or every 16 ms; thus, reactions that result in the activation of different behaviors occur quickly and often blend together in the eye of the watcher to seem integrated. Perhaps, this type of behavior is even at the core of evolved intelligence (Brooks 1999 ) (Kennedy et al. 2001 ).

The order of behaviors is critically important to the overall operation of the system. For example, if the height behavior was prioritized over the recharge behavior, the robot would never be able to charge its batteries. Every time it tried to land at the battery charging station, the height module would make it climb again! If the avoid behavior was ordered below seek, the robot would run into obstacles and likely crash while moving to its destination. Thus, the emergent intelligence of these robots is a product of the careful, simultaneous consideration of both wholes and parts (Arnold and Wade 2015 ). The desired result emerges from the determination of what each behavior should do in the context of the others and how the behaviors are correctly prioritized as a whole system.

A major advantage to this approach is flexibility in the software; in the software designed for this research, behavior modules can be coded and inserted by outside parties. A simple configuration file determines their load order (priority), and they can be added to the system by simply placing the compiled behavior module in the Behaviors folder on the host computer’s hard drive. In this way, the simulation system is extremely flexible in that it allows testing of all sorts of behaviors and orders without requiring any changes to the base system.

Method implementation

The details of each of the behaviors and control methods are explained in this section. It is important to note that UAVs are continuously broadcasting their own locations over a wireless network and receiving and processing the locations of other UAVs.

Launch— Take off from a stationary position

Activation: Robot is not flying, is within 10 m of deployment location, and has at least 99% battery life.

Begin ascending. Note that nothing more is needed; once the robot is flying, the height module will take over and bring it to the correct altitude.

Results: Robot will ascend from a previously landed position.

Avoid— Avoid collisions with buildings and obstacles

Activation: Potential collision detected based on speed, angle of movement, acceleration, and location of nearby objects as reported by sonar sensor.

If moving faster than acceleration rate, decelerate.

If moving slower than acceleration rate, accelerate full speed at a 200° angle from current heading. This essentially turns the robot in the opposite direction of the imminent collision, at a slight 20° angle difference. The 20° angle difference prevents the robot from moving straight backwards, and then forwards again into the same situation as the previously executed behavior takes over.

If, after 12 s, robot is still within 2 m of the original location, change the deflection degrees from 200 to 160 (20° angle on the other side of the opposite.

Results: Robots will “bounce around” objects in their way.

Climb— Climb over obstacles

Activation: An obstacle is closer than 5 m as detected by sonar sensor.

Accelerate upwards at maximum acceleration, until obstacle is not detected horizontally to robot.

Stabilize horizontal movement during upwards acceleration.

Results: As a robot nears an obstacle, it will ascend up over the obstacle, where the height module then takes over and brings the robot to the appropriate height above the obstacle.

Recharge— Recharge batteries

Activation: Less than 5 min of battery life left.

Move directly to deployment location at 75% of maximum speed.

If within 3 m of deployment location, reduce speed until stabilized, then land.

Results: When a robot’s battery becomes low, it flies directly back to the deployment location and lands.

Height— Maintain a certain height above the ground or large objects

Activation: Closest object below robot is six or more meters away or four or less meters away.

If closest object is six or more meters away, descend at maximum acceleration.

If closest object is four or less meters away, ascend at maximum acceleration.

Results: Robots tend to maintain the desired height above objects below them.

Spiral— Move outwards in an expanding spiral

Activation: Four or more survivors detected within a 10-m radius of each other.

Move in an expanding spiral from the center point of the located survivors until reaching a 100-m radius.

Results: This behavior can be equated to the “expanding square” visual search pattern (Washington State Department of Transportation 1997 ) but is implemented as an expanding circle instead of a square. When the UAV detects a concentrated group of survivors, it begins to spiral outwards from the center location of the survivors. As survivors often congregate in larger groups and move towards groups, it is theorized that this behavior will lead to the discovery of additional survivors that may not have been able to reach the detected group.

Form— Maintain 50 m ± 5 m distance between other robots

Activation: Closest robot is either within 45 m or more than 55 m away.

If within 45 m, accelerate in opposite direction of closest robot at maximum acceleration rate.

If more than 55 m away, accelerate towards closest robot at maximum acceleration rate.

Results: This is a type of flocking behavior (Kennedy et al. 2001 ). Robots tend to group up together and stick together in large groups. Small groups can split off, but as they move near each other, they tend to re-engage the larger group.

Repel— Stay at least 10 m away from other robots

Activation: Closest robot is within 10 m.

Accelerate in opposite direction of closest robot at maximum acceleration rate.

Results: This behavior prevents robots from moving too close to each other in the absence of a flocking behavior such as form .

Seek— Move directly to specified GPS location

Activation: Seek location specified, and robot is more than 10 m away.

Accelerate towards specified location at maximum acceleration rate.

Results: Robots can be ordered to move directly to specific locations.

Waypoint— Move towards a preset pattern of waypoints

Activation: Set of search waypoints exists.

Accelerate at maximum rate towards current waypoint.

Once waypoint is within camera detection range, broadcast completion of waypoint over wireless network and set next waypoint as current waypoint.

Results: As the UAVs act as a single entity, they “compete” to reach the next waypoint. No single UAV is in charge, and there is no “leader” UAV. Any UAV that reaches the next waypoint will send a message to all other UAVs declaring that the waypoint has been reached. Upon receipt of this message, the UAVs will begin to move to the next waypoint. Thus, as a single system, the UAVs can be assigned one set of waypoints and they will effectively explore every waypoint as a swarm. In essence, waypoints tell the swarm to ensure that some part of your swarm, any part, covers this waypoint . In the simulations used, UAVs communicated their waypoint information via Wi-Fi. Thus, delays or long distances in Wi-Fi could have an effect on the swarm’s behavior as a whole.

The waypoint search used in this research resulted in a version of a search called “parallel track” or “parallel sweep” (Washington State Department of Transportation 1997 ) performed as a swarm. Also, when this behavior combines with avoid , the UAVs perform a variation of the “contour search” (Washington State Department of Transportation 1997 ) because they automatically avoid collisions. These are some of the interesting emergent properties of the interactions between simple behaviors.

Scatter— Move towards a pre-defined search pattern waypoint which is not already allocated to another UAV

Once waypoint is within camera detection range, broadcast completion of waypoint over wireless network and set next waypoint as current waypoint. Next waypoint must not be the current waypoint of any other UAV in the system.

Results: The swarm of UAVs scatters across the disaster area, searching multiple different locations simultaneously.

Wander— Choose a random location and move towards it

Activation: Always. Note that this behavior is rarely activated in a fully functioning system because it is almost always subsumed by some other behavior.

If location sensor exists and is functioning, choose a random wander location 100 m away and accelerate towards it at half speed.

If within 10 m of current wander location, choose new location.

If a location sensor does not exist or is malfunctioning, set a random target heading and proceed at half speed.

After traveling for 1 min at current heading, change to a different heading.

Results: This behavior is included for robustness. Wander is a default behavior in case other behaviors crash or fail to execute for any reason. If all else fails, a UAV will try to wander to a new location which may have different sensory inputs and/or different terrain, facilitating a better result.

Table  1 shows the behaviors used by each control method. Although it may appear that these methods are similar in that they use many of the same behaviors, notably most of those behaviors are a necessary foundation to the successful function of any higher-order robot behavior. A living being must eat, drink, and breathe before she can do more complex tasks. In the same way, our UAVs must launch, avoid obstacles, and maintain height before they search for disaster survivors. The essential, method-defining behaviors are the ones included, or left out of, each method.

Standard method

A swarm of UAVs operating the standard method behavior set (Fig.  2 ) will launch then proceed to the first waypoint in their search pattern (Fig.  3 ). Along the way, they will maintain appropriate distances between each other by continuously broadcasting their locations over a wireless network, avoid collisions with obstacles through maneuvering around or climbing over, and maintain proper height. When the first UAV in the swarm reaches the current waypoint location, it broadcasts this data to the rest of the swarm. As the UAVs receive this data, they begin moving towards the next waypoint in the search pattern. In some cases, UAVs on the far side of the swarm may already be close to the new waypoint. The result is that a large swarm of UAVs may “zig-zag” between locations in a way that can be efficient, whereas a smaller swarm of just one, two, or three UAVs may actually fly back and forth between the waypoints. Both methods maximize coverage area and follow the same behavior software, though an observer will notice significant differences in the actual flight paths of the UAVs and may conclude (incorrectly) that they are actually using different artificial intelligence software.

Standard method behavior set

Standard method showing the paths of three UAVs launched from the blue rectangle on the center left. Red, yellow, and green dots are survivors in different states of discovery. In this scenario, UAVS moved in a search pattern across the area starting in the northwest and ending in the southeast. Photograph by Geospatial Information Authority (GSI) of Japan (Geospatial Information Authority of Japan 2011 )

Upon a low battery indication, a UAV will break from formation and return to its deployment location, land, and recharge its batteries. When the recharge is complete, the launch behavior will detect a full battery and automatically activate. The robot will then launch and proceed to the next waypoint, likely meeting up with the rest of the swarm along the way.

While following this method, it is possible and likely that robots will break into smaller groups as they recharge their batteries and return to the field. The design and architecture do not prevent or discourage this, and it is an emergent result of the complex interactions of simple behaviors.

Spiral method

The spiral method uses the standard method but implements an additional behavior: spiral, which is inserted after height and before form in the behavior priority list.

The spiral method behavior set (Fig.  4 ) operates similarly to the standard method, but differs in one significant way. While engaging in the standard method search, when a UAV’s spiral behavior is activated through detection of a concentration of survivors, the UAV “breaks away” from the group and performs a spiral maneuver out to a 100-m radius (Fig.  5 ). After completing this maneuver, the robot returns to its regular formation within the group. Within the software architecture, the only requirement to implement this method is the insertion of the spiral behavior module in the correct place in the behavior list. No other changes need to be made. That such a change can be made so simply is one of the advantages of the behavior-based artificial intelligence paradigm.

Spiral method behavior set

Spiral method showing the paths of three UAVs. As with standard method, UAVs launched from the blue rectangle. Note the circular pattern in the northeast corner as a UAV located the group of survivors (green dots) on top of the elevated building and performed the spiral behavior while the others continued the search. Photograph by GSI of Japan (Geospatial Information Authority of Japan 2011 )

The spiral method accounts for evidence gathered during disaster search and rescue (Editorial Office of the Ishinomaki Kahoku 2014 ) (A. E. S. M. Staff Member 2017 ) showing that survivors are likely to group together following a disaster. If a few people are found together, it is likely that more are present as well. Spiraling outwards from the locations of the first few people found is likely to result in the discovery of new survivors.

The distressed person density information could be used by rescue workers in many ways, such as determining where and when to send rescue vehicles such as helicopters or boats. Also, the spiral method may result in the discovery of distressed persons attempting to unite with the group, and coming close, but failing to cover the last bit of distance due to insurmountable obstacles, as happened during the 2011 tsunami (Editorial Office of the Ishinomaki Kahoku 2014 ).

Scatter method

The scatter method differs from standard and spiral methods significantly in that it sends each UAV to a different point in the search pattern. The waypoint behavior module is removed completely and replaced with a scatter module. Also, the form module is replaced with the repel module.

The scatter method (Fig.  6 ) represents a significant diversion from both the standard and spiral methods. Although this method is still cooperative, rather than operating as single flock with all robots seeking the same point then switching to the next when any one UAV reaches the point, using the scatter method, each UAV has its own destination point which is different from all the others (Fig.  7 ). Theoretically, this allows the swarm to spread over a larger area in shorter time.

Scatter method behavior set

Scatter method showing the paths of three UAVs. As with standard method, robots launched from the blue rectangle. However, each UAV proceeded to a different location in the search pattern, scattering them across the area. Photograph by GSI of Japan (Geospatial Information Authority of Japan 2011 )

Destinations are selected based on a staleness factor , that is, points that have not been reached yet by the swarm as a whole are highest priority, whereas points that have been visited further in the past are slightly lower, and points that have been recently visited are the lowest in priority. If one UAV is already seeking a point, a different point is chosen. If all points are already chosen, the UAV chooses an optimal point based on staleness factor. Using this method, the swarm of UAVs will effectively scatter across the disaster area, searching multiple different points simultaneously.

Although in theory the scatter method might appear to be a better option than standard or spiral methods given that different UAVs are able to explore different locations in parallel, in practice, a swarm of UAVs flocking together significantly increases the probability of survivor detection. Sensor range is limited, and a group of UAVs flocked together maintaining a certain distance from each other effectively forms a large, single system with a combined, redundant sensor range. Without flocking, a single UAV’s sensor range is limited; therefore, as locations are explored separately, the search pattern must necessarily be quite complex or contain a large number of waypoints to approach the same level of effectiveness as the other methods. In this case, a hybrid method between scatter and spiral could be more effective.

Performance analysis

Assumptions.

While developing the simulation software used in this research, several assumptions were made about the UAVs:

Programmable —The UAVs are programmable in that they are controlled by modifiable software and can receive commands to change speed and direction.

Quadcopter —UAVs are standard multirotor helicopters lifted and propelled by four rotors.

Stability control —UAVs have built-in stability control that allows them to hover stably in one location or can be easily equipped with equivalent Commercial Off the Shelf (COTS) software to provide this effect.

Network unavailable —Due to loss of infrastructure and other inherently challenging circumstances during most disaster search and rescue situations, it is assumed that a commercial Internet network may not be available. The UAVs will set up their own ad hoc network to communicate with each other. This network is not dependent on existing network infrastructure.

The simulation software allows the selection of different commonly available off-the-shelf UAVs. It also allows UAV parameters to be customized. For the scenarios used in this research, Table  2 shows the parameters that were used in the simulation based on current commercially available data.

Sensors and equipment

In addition to the software behavior modules, UAVs are provided with simulated sensors and equipment values to be customized (Table  3 ). Collision avoidance depends on sonar sensors. One sonar sensor is mounted down-facing, while the others are outward-facing from the left, right, forward, and rear sides of the UAV. The sonar data is fused together to form a single sonar sensor picture. Formation and flocking behavior depends on both sonar sensors and the GPS. Communication between UAVs, and therefore cooperative swarm behavior, depends on the Wi-Fi HD communicator. UAVs determine their own locations, and, by extension, which direction to travel to reach a waypoint, by using the GPS sensor. The behavior modules are highly dependent on the input from these sensors.

These sensors can be turned off or on, or “broken” in the simulation to simulate how a UAV will behave in different practical situations. The range and effectiveness of the sensors can also be adjusted. This allows the designing of a robust system prior to actual deployment and hardware testing.

The UAV’s camera is mounted in a down-facing position on the bottom of the chassis. Although a camera radius of just 15 m may seem small, the intent of this range is to capture difficult environmental conditions such as fog, snow, rain, and debris, which may interfere with a camera’s range of vision. A 15-m radius provides a conservative estimate that likely falls within the effective parameters of a wide range of commercially available cameras and sensors.

Simulation scenario

The environment chosen to be simulated for this research was in a town called Arahama, in Wakabayashi, Sendai City, Miyagi Prefecture, Japan, one day after the 2011 Great Eastern Japan Earthquake and Tsunami. This location was chosen because it was one of the hardest hit by the tsunami, and a great deal of data were available on the town, including satellite imagery, population, physical layout, timetable of the tsunami, search and rescue data, personal interviews, and locations of survivors. Within this environment, three different patterns were considered when setting the locations of distressed persons within the simulation (Fig.  8 ):

Random —Distressed persons were scattered at random across the search area.

Congregated —Distressed persons were concentrated at likely rescue locations according to data from a variety of sources. For example, schools, parking decks, and other tall buildings contained more survivors while low areas contained few, if any (Editorial Office of the Ishinomaki Kahoku 2014 ) (Municipal Development Policy Bureau 2017 ) (Post-Disaster Reconstruction Bureau 2015 ) (Tohoku Regional Development Association n.d. ) (A. E. S. M. Staff Member 2017 ).

Mixed —Half of the distressed persons were congregated and the other half random.

Survivor distribution patterns. Gray boxes are buildings, red dots are survivors. From left to right: random, congregated, mixed

The mixed pattern was selected and used for our research. Although the congregation pattern is based on real data acquired at Arahama (Editorial Office of the Ishinomaki Kahoku 2014 ) (Municipal Development Policy Bureau 2017 ) (Post-Disaster Reconstruction Bureau 2015 ) (The Center for Remembering 3.11 2015 ) (Tohoku Regional Development Association n.d. ) (A. E. S. M. Staff Member 2017 ), randomly scattered survivors should not be discounted as it could be that they were simply not found during rescue efforts. Therefore, the mixed pattern is the best fit for this research. Practical algorithms should show greater effectiveness at congregation-heavy patterns than at random patterns.

For the purpose of the simulation, satellite imagery of the actual location was acquired (Figs.  9 and 10 ). Ideally, a photo immediately following the tsunami strike would be desirable. Unfortunately, such imagery was not available; this image was taken on March 12, 2011, the day after the tsunami strike. Building locations were placed according to the imagery and checked against height data as well as cross-referenced against actual photos and on-site interviews with local residents.

Satellite photo of the town of Arahama taken on March 12, 2011 (Geospatial Information Authority of Japan 2011 ). Left is full photo, and right is sample of 300 m 2 sub-section built in DroneLab environment builder showing red buildings and red dot survivors. The large buildings in the upper left corner of the right photo are the ruins of the Sendai Arahama Elementary School, a primary evacuation site during the tsunami. Photograph by GSI of Japan (Geospatial Information Authority of Japan 2011 )

Left is northwest corner of satellite photo in Fig.  9 , and right is the same area after build-out using DroneLab environment builder. Buildings shown as red rectangles and survivors as red dots. Photograph by GSI of Japan (Geospatial Information Authority of Japan 2011 )

Validation of the disaster area model

The model of the disaster area was built by overlaying structural data on the satellite photos shown above, resulting in a high level of face validity. The heights of the buildings were determined by on-site survey and measurement. As time did not allow for all buildings to be measured and some have in fact been demolished since 2011, buildings that could not be directly measured were assigned height data based on their types, locations, and designs. For example, in a row of similar houses, the height of a single house may have been measured and then used for all similar houses.

To accurately represent survivor distribution in the simulation model, data from a variety of sources were used. These data can be collated to show a pattern in which groups of certain numbers of survivors gathered at certain places within the town (Editorial Office of the Ishinomaki Kahoku 2014 ) (Municipal Development Policy Bureau 2017 ) (Post-Disaster Reconstruction Bureau 2015 ) (The Center for Remembering 3.11 2015 ) (Tohoku Regional Development Association n.d. ) (A. E. S. M. Staff Member 2017 ).

As no data is available on the locations of victims lost to the tsunami in Arahama, a random distribution pattern was chosen to represent the remainder of the town’s population. The mixed pattern using the real data combined with the use of random distribution for the remaining survivors, based on the total population of the town, is considered a reasonable way to represent the survivor locations in the simulation model based on available data.

The model was validated by comparing the locations and heights of buildings, numbers of survivors, and congregated groups of survivors to satellite photos, aerial photos, and records obtained in Arahama detailing the events during and immediately following the tsunami. The resultant simulation model was used as a base for the simulations performed during this research.

Results in simulation

Results were generated using the DroneLab Unmanned Aerial System (UAS) simulation software sponsored by the Japan Acquisition, Technology, and Logistics Agency’s (ATLA) Air Systems Research Center (ASRC). Footnote 5

DroneLab runs on multiple platforms, including macOS, Unix-like operating systems, or Microsoft Windows machines, using the Java environment. The simulation environment is user-definable, displaying either an image as a background or a blank field of 2000 × 2000 m. A background image is typically a satellite photo of arbitrary size. The environment is three dimensional, displaying both a two-dimensional top-down view and a three-dimensional view. Various sizes, heights, and dimensions of square, circular, and rectangular objects can be placed on the field both before and during a simulation. Survivors can also be placed on the field at specific locations and/or distributed randomly. Deployment locations for rescue workers can be placed as rectangular areas on the field. The aerial robots are displayed as circles with spinning bars in their centers, whereas the obstacles are red objects in the two-dimensional view and yellow objects in the three-dimensional view. Survivors are shown as red dots on the field, turning yellow and finally green based on their states of discovery. Sensor range displays can be toggled on and off from the simulator’s user interface.

DroneLab allows the inclusion of one or many robots equipped with simulated sensors and equipment and supports the addition of pluggable behavior modules written in the Java programming language. It includes a physics engine that allows specification of speeds, acceleration rates, and various other physical properties, and provides collision checking and gravity. DroneLab allows the acceleration of time and the addition of obstacles “on-the-fly” to create a dynamic virtual environment.

Figures  11 and 12 show the results of the simulation when applied to swarms of UAVs using the parameters presented in Table  2 . The percentage of survivors seen over time by the IR camera, referred to as camera coverage , was chosen as the measure to display, as its range can be generalized to many other sensors. Each UAV’s simulated camera’s detection radius was limited to 15 m as a way to account for environmental conditions such as darkness, fog, rain, snow, and debris. The camera coverage percentage shown in the vertical axis of the figures is a measure of the number of total survivors detected by the camera of any UAV divided by the total number of survivors in the simulation, the ratio of detected to undetected survivors by the swarm as a whole. Survivor distributions use the mixed method described previously in Fig.  8 . Three hundred fifty survivors were congregated on and around likely evacuation sites (Editorial Office of the Ishinomaki Kahoku 2014 ) and 300 were scattered randomly across the disaster area, for a total of 650 survivors. According to sources from Arahama (Editorial Office of the Ishinomaki Kahoku 2014 ) (Sato 2015 ) (A. E. S. M. Staff Member 2017 ), the number 650 is roughly equal to the population of the local area at the time of the tsunami. The time axis shows simply the hours, minutes, and seconds since the UAV swarm was deployed.

Average percentage of survivors found over time, referred to as camera coverage, by a swarm of five UAVs across six simulation runs. Scatter method was the slowest and spiral method the fastest to reach the goal of 90%. The jump in coverage just after 43 min occurs when the swarm encounters an evacuation center such as a school in which many survivors are co-located

Average camera coverage rates of swarms of five, 10, and 20 UAVs for all three methods shown in parallel. ST is standard , SP is spiral , and SC is scatter . Scatter method with five UAVs was the slowest and spiral method with 20 UAVs the fastest to reach the 90% coverage goal

Figures  11 and 12 show that in all of these results, every situation resulted in the achievement of 90% or more camera coverage in under 2 h. Swarms of 10 or 20 UAVs using the standard or spiral method were consistently able to discover 90% or more of simulated survivors in less than an hour. Left running for 4 h, swarms of 10 or more UAVs consistently achieved a 98% or 99% location rate as they re-ran their routes in flocking formation. Similar algorithm differences were observed when the UAVs were equipped with simulated 5-m-ranged FINDER sensor instead of the camera, though discovery times generally increased by 30–50% with the shorter-ranged sensor.

These results are significant as they show that there is the potential to spot 90% of visible survivors of a disaster situation, even in hazardous, non-drivable, or inundated areas, in under an hour with little operator intervention using the proposed technique. This is well within the 24-h time limit suggested as optimal for disaster response (Editorial Office of the Ishinomaki Kahoku 2014 ) (Bartels et al. n.d. ), even when the potential multi-hour mobilization times for manned rescue teams are factored in. These results almost certainly represent a significant improvement over existing methods. Actual data regarding the time it takes rescue workers to thoroughly search an equivalent area using existing methods without the use of UAVs varies by situation and is difficult to quantify. However, available evidence suggests that it can take days to search the most significantly affected areas (Editorial Office of the Ishinomaki Kahoku 2014 ).

Additionally, there are many situations in which long-term search and rescue efforts are necessary and difficult to sustain by manned personnel (American Red Cross 2015 ). At times, survivors are discovered days or even weeks after the initial disaster strike. In these situations, swarms of UAVs may continue operating and searching with little human interaction needed to achieve a high degree of sensor coverage over a short period of time. A swarm of 10 UAVs using the spiral method on average was able to achieve 98.9% camera coverage in under 90 min. This rate slowly grows over time due to the unpredictable nature of the swarm patterns. Each time the pattern is re-flown, the positions of each UAV differ due to responsive flocking behavior. This element of randomness improved long-term search results and could be leveraged to a higher degree in a non-simulated system.

Despite these results, it is important to acknowledge that in order for this data to be useful at present, a human rescue worker would necessarily view and process the data so that survivors could actually be rescued. The swarms of UAVs simulated in this research are not intended to perform actual rescues, although such efforts are possible (Erdelj et al. 2017 ) (American Red Cross 2015 ). Therefore, this research acts as an initial step to demonstrate what lies within the realm of the possible using a behavior-based UAV swarming approach to disaster search and rescue. The research also provides suggestions for initial algorithms and search methods that have proven effective in simulation.

Areas of improvement

Despite the positive outcome evident in the simulation results, one persistent cause of delay across all scenarios was the battery recharge behavior. The flattening discovery rates in each simulation run are often caused when UAVs run out of batteries around the 20–25 min time-frame and return to the deployment location for a recharge. This happens repeatedly as batteries discharge, sometimes requiring a UAV to fly across the entire area to return to the charging station. Intelligent recharging to handle this behavior could improve results of the algorithms further. For example, future iterations of the search algorithms could integrate battery recharge into their designs, potentially triggering an auto-charge when a UAV flies within a certain distance of its recharge station while its battery life is below a certain threshold.

Applications

This section provides a sample of practical applications for which this research can be leveraged, as well as brief guidance on how to apply the research to actual situations. Although some aspects of this research are experimental, such as the simulated miniaturized FINDER sensor, other aspects, such as the use of a UAV swarm equipped with Wi-Fi and IR cameras, are readily usable today.

At present, the recommended deployment configuration is 5 or 10 commercially available UAVs with parameters as good as or better than those specified in Table  2 . Each UAV should be equipped with an infrared (IR) camera and loaded with the software used in this research. Additional work would be necessary to be done to pull data from the real, rather than simulated, IR cameras and other sensors.

To accept the data feeds from the UAV swarm, a ground station and/or mobile application could be developed. This application could be designed in many different ways, but the basics could include a top-down graphical map, photo, or blank image of the search environment with a built-in customizable distance scale such as the one used in the DroneLab simulation software. As survivors are located, the operator or operators could tap the screen to indicate their locations. At present, the recognition of humans from camera feeds is a challenging research problem in and of itself. Thus, the rescue personnel could view the data feeds of the various UAVs and mark locations on a shared map. The combination of cooperative UAV swarm, mobile application, and input from rescue personnel would form a viable mode of operation using technology available today.

Types of disasters

The approach described in this research is well suited to earthquake and tsunami disasters, as well as any destructive natural or man-made disasters in which environmental or political conditions present difficulty in the deployment of rescue vehicles or personnel. These situations include the presence of significant or hazardous debris, inundated terrain, and/or dangerous or hostile conditions. Due to limits in UAV communication range and battery life, the cooperative behavior is optimized over a 2-km 2 area. Thus, the approach is particularly well suited to environments in which the presence of undiscovered survivors within a particular area is strongly suspected—for example, within cities, towns, villages, or other populated areas.

Civil/defense applications

In addition to civilian search and rescue, this research also has a number of applications that could apply to both the civil and defense sectors. With much detail omitted, the following are a list of potential applications in which the use of swarms of autonomous cooperating UAVs such as those simulated in this research could be highly valuable:

Intelligence gathering

Combat search and rescue

Smart object location acquisition

Incoming threat detection

Site assessment and map-building

Counter-UAS and counter-swarming

Ethical considerations

A number of ethical considerations surround the use of humanitarian robotics. One such consideration is the fact that swarming algorithms and autonomous robotic systems in general are inherently dual-use. These systems can often be used for civilian or military purposes. Although this research focuses on the use of UAV swarms for humanitarian disaster relief and the defense applications outlined above do not specifically recommend weaponization of this research or technology, such an outcome is possible.

The use of fully autonomous systems in weapons systems opens the potential for a new type of risk. When implemented on weapons platforms, autonomous systems can select and attack targets in ways that are faster and different than those performed by humans. Due to the potential for unintended collateral damage caused by these systems, the United States Department of Defense does not permit lethal fully autonomous weapons systems at this time (Human Rights Watch 2013 ). All weapons that include artificial intelligence must also include a human supervisor, or “human-in-the-loop,” for decision-making (US Department of Defense 2012 ).

In addition to its dual-use nature, other ethical concerns are inherent to humanitarian robotics research. Any time a machine is empowered with the ability to make or influence decisions that affect peoples’ lives, ethics becomes an important factor in system development and deployment (Sandvik et al. 2014 ). When designing a system based on this research, these factors should be among those considered as part of a comprehensive systemic ethics policy.

If an autonomous robot swarm is used to detect and report the locations of survivors, what issues might cause bias in reporting? Computer algorithms are developed by humans and cannot be said to be entirely free of bias and politics (Sandvik et al. 2014 ). Different algorithms, test cases, or detection equipment could create bias in the detection and reporting process.

Certain people or types of people may be reported over others. For example, automated face recognition techniques tend to be more effective on certain ethnic groups (Sharkey 2018 ). If such techniques are used by the swarm system to detect survivors, there is likely to be detection bias.

As a behavior-based approach creating emergent intelligence, how might ethics be examined differently in the case of this research than it would be in a centrally controlled system?

Do behavior-based artificial intelligence systems fall under the same sets of considerations as centrally controlled systems?

Another valid ethical concern in humanitarian robotics is the issue of neutrality. Neutrality can be compromised if UAVs are perceived—even if incorrectly—to be linked to a military or political power that has a stake in a humanitarian crisis (Emery 2016 ). Engagement with the local community is one way to approach this concern. However, with regard to the research described in this article, the UAVs used for this research are commercial or custom quadcopters commonly used by drone hobbyists. These UAV models are not likely to cause tension or misperception as might repurposed military UAVs.

Perceptions of the ethical issues surrounding UAVs also differ in different parts of the world. For example, in Europe and North America, concerns about the use of UAVs tend to include invasion of privacy, misuse by government or law enforcement, and fears of an aviation disaster. However, concerns in the Tana Delta of Kenya, where humanitarian drones were field tested, revolved around practical concerns such as the strength of the UAV’s camera, how far the system could operate, how quickly the drones could be deployed in an emergency, and who would be in physical possession of the system (Boyd 2014 ). Given this knowledge, it is important to consider the concerns of the local communities with regard to humanitarian drones, rather than to superimpose the concerns of aid-providing nations in the mistaken assumption that the concerns are identical.

The results of the study appear to greatly improve the availability of situational awareness data in the first few hours after a major natural disaster, which is widely considered one of the most critical SAR areas in need of improvement (Editorial Office of the Ishinomaki Kahoku 2014 ) (Erdelj et al. 2017 ) (Tait Communications 2012 ) (Bartels et al. n.d. ) (Ochoa and Santos 2015 ) (Shimanski 2005 ) (Adams et al. 2007 ) (Riley and Endsley 2004 ). Simulation data generated during the study show that a swarm of just five standard Footnote 6 UAVs executing the spiral method of cooperative, behavior-based search and rescue developed in this research can consistently achieve a 98.8% 15-m radius sensor coverage after 4 h, reaching a goal coverage rate of 90% in 90 min. This same swarm of five UAVs consistently achieves a 92.5% 5-m radius sensor coverage rate in 4 h, reaching the 90% goal in 3 h. As more robots are added, the numbers improve even more. A 10-UAV swarm averages 98.9% standard sensor coverage after 4 h and reaches a 90% coverage in only 53 min. Equipped the precise 5-m radius sensor, 10 UAVs reach 96.9% coverage after 4 h, reaching the goal 90% in 108 min.

In many simulations, a swarm of 20 UAVs using the spiral method reached the 90% goal in less than 34 min, slightly over half an hour to discover 90% of all visible survivors within a 2-km 2 area littered with waterlogged fields, damaged structures, fallen trees, and overturned piles of cars.

The spiral method is likely the quickest because it reacts more effectively to groups of survivors. The spiral method discovers clusters of survivors more quickly than the other methods through its spiral behavior module, which spirals outwards from an area in which more than a certain number of survivors are detected. If different types of data were sought after, a different set of behaviors might prove more effective.

Given the strong results of the simulations performed as part of this research, this approach to post-disaster assessment appears promising. Of course, in a real-world situation, the usage and availability of the data discovered by the UAVs is key. Also, although these simulations were designed to model a real environment with some degree of accuracy, the performance will certainly differ in an actual situation. However, this research does show that the use of swarms of UAVs with these algorithms has the potential to make a large amount of critical data available for consumption by rescue workers or other systems of interest. This research demonstrates the potential for high value in the area of disaster data acquisition using swarms of autonomous UAVs.

Change history

01 june 2019.

Following publication of the original article [1], the authors reported errors.

Standard parameters such as those of the commercially available DJI Phantom 4 quadcopter or similar model.

Standard sensor coverage for this research is considered to be a 15-m radius detection range.

The UK’s ORCHID Project seeks to create a disaster response system using a swarm of UAVs at a cost of around $2000 each (Kube and Zhang 1992 ).

A Finding Individuals for Disaster and Emergency Response (FINDER) sensor is a sensor developed by the US National Aeronautics and Space Administration (NASA) to aid in disaster search and rescue. A FINDER sensor uses low-power microwaves to detect the heartbeats of buried disaster survivors up to 9 m into a mound of rubble. It has been used to successfully locate survivors in Nepal. A FINDER sensor is currently the size of a carry-on bag and is thus not appropriate for carry by a standard quadcopter. However, simulating how a future miniaturized version of this sensor, or others like it, might perform alongside a standard visual or infrared camera provides an interesting comparison for the purposes of this research.

The DroneLab simulation software, as well as the UAV controlling software, may be available upon request to the (Institution omitted for blind paper submission) or through request to the paper’s author. At the time of this writing, the software is not public domain.

A. E. S. M. Staff Member, Interviewee, Description of events at Arahama during the Tohoku tsunami. [Interview]. 2017.

Adams AL, Schmidt TA, Newgard CD, Federiuk CS, Christie M, Scorvo S, DeFreest M (2007) Search is a time-critical event: when search and rescue missions may become futile. Wilderness and Environmental Medicine 18(2):95–101.

Article   Google Scholar  

Alley RE (1992) Problems of search and rescue in disasters. In The Management of Mass Burn Casualties and Fire Disasters. Dordrecht, Springer Netherlands. pp. 175–176. http://doi.org/10.1007/978-0-585-33973-3_2. .

American Red Cross, “Drones for disaster response and relief operations,” 2015.

Google Scholar  

Arnold RD, Wade JP (2015) A definition of systems thinking: a systems approach. Procedia Computer Science 44:669–678.

Balch T, Arkin RC (1998) Behavior-based formation control for multi-robot teams. IEEE Transactions on Robotics and Automation 14(6):926–939. http://doi.org/10.1109/70.736776 .

Bartels R, Herskovic V, Monares A, Ochoa SF, Pino JA, Roges MR (2010) A simple and portable command post to coordinate search and rescue activities in disaster relief efforts. In: 16th international conference on collaboration and technology. CRIWG, Maastricht.

Boyd D, “Humanitarian drones: perceptions vs. reality in Kenya’s Tana Delta, The Sentinel Project , 2014.

Brooks RA (1999) Cambrian intelligence: the early history of the new AI. The MIT Press, Cambridge, MA.

Editorial Office of the Ishinomaki Kahoku (2014) Surviving the 2011 tsunami: 100 testimonies of Ishinomaki area survivors of the great East Japan earthquake. Junposha Co., Ltd, Mejirodai, Bunkyo-ku, Tokyo.

Emery JR (2016) The possibilities and pitfalls of humanitarian drones. Ethics and International Affairs 30(2):153–165.

Erdelj M, Natalizio E, Chowdhury KRAIF (2017) Help from the sky: leveraging UAVs for disaster management. In: Pervasive Computing, pp. 24–32, January–march.

Frelinger D, J. Kvitky and W. Stanley, Proliferated autonomous weapons: an example of cooperative behavior, RAND Corporation, 1998.

Geospatial Information Authority of Japan, Arahama, Sendai on March 12, 2011, 2011.

Guarnieri M, Kurazume R, Masuda H, Inoh T, Takita K, Debenest P, Hodoshima R, Fukushima E, Hirose S (2009) HELIOS system: a team of tracked robots for special urban search and rescue operations. In: IEEE/RSJ international conference on intelligent robots and systems. IEEE, St. Louis.

Hambling D, Watch 50 drones controlled at once in record-breaking swarm, New Scientist , 2015.

Human Rights Watch, Review of the 2012 US policy on autonomy in weapons systems,2013.

Kennedy J, Eberhart RC, Shi Y (2001) Swarm intelligence. Morgan Kaufmann Publishers, San Francisco, CA.

Koffka K (1922) Perception: an introduction to the Gestalt-Theorie. Psychol Bull 19:531–585.

Kube CR, Zhang H (1992) Collective robotic intelligence. In: Second international conference on simulation and adaptive behavior, Honolulu. MIT Press Cambridge, MA, USA.

Macintyre A. G, J. A. Barbera, E.R. Smith, Surviving collapsed structure entrapment after earthquakes: a time-to-rescue analysis, Prehospital and Disaster Medicine , vol. 21, ch 1, pp. 4–19, 2006.

Molina P, Pares ME, Colomina I, Vitoria T, Silva PF, Skaloud J, Kornus W, Prades R, Aguilera C (2012) Drones to the rescue! Unmanned aerial search missions based on thermal imaging and reliable navigation. In: InsideGNSS, pp. 38–47, July–august.

Municipal Development Policy Bureau (2017) Ruins of the great East Japan earthquake: Sendai Arahama elementary school. United Nations Office for Disaster Risk Reduction, Sendai City.

Ochoa S, and R. Santos, Human-centric wireless sensor networks to improve information availability during urban search and rescue activities, Information Fusion, pp. 71–84, 2015.

Pinciroli C, & Beltrame G. (2016). Buzz: An extensible programming language for heterogeneous swarm robotics. In 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 3794–3800.IEEE. https://doi.org/10.1109/IROS.2016.7759558

Post-Disaster Reconstruction Bureau (2015) Reconstruction of Sendai. In: Third UN conference on disaster risk reduction. United Nations Office for Disaster Risk Reduction, Sendai City.

Ramchurn SD, Wu F, Fischer JE, Reece S, Jiang W, Roberts SJ, Rodden T, Jennings NR (2016) Human-agent collaboration for disaster response. Journal of Autonomous Agents and Multi-Agent Systems 30(1):82–111. http://doi.org/10.1007/s10458-015-9286-4 .

Riley J. M, M. R. Endsley, The hunt for situational awareness: human-robot interaction in search and rescue, in Proceedings of the Human Factors and Ergonomics Society Annual Meeting , 2004.

Sandvik K.B, et al. Humanitarian technology: a critical research agenda. International Review of the Red Cross 96.893, 2014, pp. 219–242.

Sato Y (2015) Museums and the great East Japan earthquake. Sendai Miyagi Museum Alliance, Sendai City.

Sharkey N, The impact of gender and race bias in AI, Humanitarian Law and Policy , 2018.

Shimanski C, Situational awareness in search and rescue operations, in International Technical Rescue Symposium , 2005.

Tait Communications, Race against time: emergency response - preventing escalating chaos in a disaster, Tait limited, 2012.

The Center for Remembering 3.11 (2015) Activity report of the center for remembering 3.11. In: Third UN world conference on disaster risk reduction, Sendai City.

Tohoku Regional Development Association (2015) Tohoku regional development association earthquake disaster response: march 11th, 2011 the great East Japan earthquake. In: Third UN world conference on disaster risk reduction. Sendai City, Japan.

US Department of Defense, Autonomy in weapons systems, Directive Number 3000.09, 2012.

Vásárhelyi G, Virágh C, Somorjai G, Tarcai N, Szörényi T, Nepusz T, Vicsek T (2014) Outdoor flocking and formation flight with autonomous aerial robots. In: IEEE/RSJ international conference on intelligent robots and systems. IEEE, Chicago.

Virágh C, Vásárhelyi G, Tarcai N, Szörényi T, Somorjai G, Nepusz T, Vicsek T (2014) Flocking algorithm for autonomous flying robots. Bioinspiration & biomimetics 9(2):025012.

Washington State Department of Transportation (1997) Visual search patterns, pp 177–191.

Wiener N (1948) Cybernetics: or control and communication in the animal and machine. The MIT Press, Cambridge, MA.

Williams M, Researchers envisage swarms of tiny drones for dangerous rescue missions, PCWorld ,2015.

Download references

Acknowledgements

Not applicable.

Funding was provided by the Japan Acquisition, Logistics and Technology Agency and the United States Department of Defense. The authors are employees of these agencies.

Availability of data and materials

Please contact the author for data requests.

Author information

Authors and affiliations.

United States Army Armament, Research, Development, and Engineering Center (US Army ARDEC), United States Department of Defense, Picatinny, NJ, USA

Ross D. Arnold

Acquisition Logistics and Technology Agency (ATLA), Japan Ministry of Defense, Tokyo, Japan

Hiroyuki Yamaguchi & Toshiyuki Tanaka

You can also search for this author in PubMed   Google Scholar

Contributions

RA carried out the research, wrote the research software, and wrote the manuscript. HY conceived of the study and participated in the research reviews and coordination. TT advised the research, participated in the design of the study, provided the background materials for the research, and participated in the research reviews and coordination. All authors read and approved the final manuscript.

Corresponding author

Correspondence to Ross D. Arnold .

Ethics declarations

Competing interests.

The authors declare that they have no competing interests.

Publisher’s Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License ( http://creativecommons.org/licenses/by/4.0/ ), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.

Reprints and permissions

About this article

Cite this article.

Arnold, R.D., Yamaguchi, H. & Tanaka, T. Search and rescue with autonomous flying robots through behavior-based cooperative intelligence. Int J Humanitarian Action 3 , 18 (2018). https://doi.org/10.1186/s41018-018-0045-4

Download citation

Received : 07 August 2018

Accepted : 10 October 2018

Published : 05 December 2018

DOI : https://doi.org/10.1186/s41018-018-0045-4

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Algorithm design behavior-based artificial intelligence
• Disaster recovery
• Drone swarms
• Multi-robot systems
• Post-disaster assessment
• Rescue robots
• Search and rescue
• Swarm intelligence
• Unmanned autonomous vehicles

The burgeoning science of search and rescue

By analyzing reports of people who got off-track, researchers are advancing the science of “lost person behavior.”

By Sarah Scoles/Undark | Published Jan 22, 2024 9:52 AM EST

• Environment

This article was originally featured on Undark .

On May 5, 2023, a 19-year-old hiker named Matthew Read headed out on a roughly 12-mile trek in an underpopulated part of Glacier National Park in Montana. Read, a chemical engineering student, had stopped in Glacier while driving home to Michigan, and the pine-surrounded path he embarked on was known for its big views of the Livingston Range, a set of jagged peaks to the east.

He would get a longer look at them than he anticipated.

By Sunday, two days later, the young hiker had yet to return. That afternoon, national park rangers started a ground search; that night, a helicopter team scanned from above. The rangers, though, were thwarted by a lack of clues, the chopper by meteorology: Clouds hung low, fog obscured the view.

That Monday, the search team grew to 30 people, and included the U.S. Border Patrol and Flathead County Sheriff’s Office, along with search dogs. It would expand to the North Valley and Flathead search and rescue, or SAR, teams.

In situations like this, a subfield of science can help those SAR teams know where to start—and how a college kid lost in big bear country might behave. It’s called, appropriately, lost person behavior.

The study expanded in the 2000s, with a researcher named Robert Koester. In 2008, Koester compiled, analyzed, and published data on how different types of people behave when wandering the wilderness—and how to find them. His work has become foundational to the field of lost person behavior, and a cornerstone of how SAR teams plan missions to find people who have wandered off the beaten path.

While it’s hard to come by solid numbers because there is no mandatory centralized place where SAR teams must file reports, in 2021, nearly 3,400 people needed help getting out of the wilderness in U.S. national parks alone (a minority of the land where people need rescue).

In a field where even minutes matter, efficient search tactics can mean the difference between life and death.

In the 16 years since Koester’s initial work, he and other researchers have added nuance, filled in gaps in the initial framework, and built new technological tools. But moving new research out of the ivory tower and into the outback isn’t simple: When searchers are seeking a kid who wandered from camp in mountain lion country or an alpinist unconscious in an avalanche, trying a new tactic isn’t often at the top of the priority list.

But in a field where even minutes matter, efficient search tactics can mean the difference between life and death. Those high stakes inspire researchers to travel farther down the path, following the data where it forks.

Systematic searches largely trace back to World War II, when spotter planes would use grid searching to detect German U-boats.

Around the late 20th century, SAR professionals in the U.S., the U.K., Canada, and Australia were formally reporting land-based rescue missions, pulling out statistics on things like why people had become lost, and for how long. The most influential of these professionals was a researcher named William Syrotuck, who analyzed a few hundred cases total, wrote up the results in book form, and represented the first attempt to sort lost people into categories (children, hunters, hikers, elderly persons), take stock of their collective actions separately, and chart how far they tended to wander.

The statistics on lost people didn’t come from large sample sizes, and so their behavior profiles weren’t quite cemented until Koester came along.

Koester’s interest in search and rescue began as a Boy Scout, when his troop did a mock emergency scenario: They were tasked with finding a man who had gone into the woods to chop down a tree and hadn’t returned. The young Koester was put in charge of strategy.

Years later, as a student at the University of Virginia, Koester joined the school’s SAR team — which was often sent out to seek lost people with dementia. It didn’t take long for Koester, who has the demeanor of a helpful park ranger, to see patterns in their behavior and connect their waypoints. Early on, he came up with a rule of thumb—literally: If he placed the bottom of his thumb where a person had last been seen, they could generally be found within the finger’s length on a standard topographic map.

One late night, his phone rang. “I get this panicky phone call from the search planner who’s like, ‘Bob, I absolutely have to know the length of your thumb,” Koester recalled.

“That’s a silly way of doing business,” he thought. Too subjective. What if, instead, he recreated Syrotuck’s statistical work, but added a new category? Syrotuck, see, had grouped everyone over the age of 65 together. But seniors with dementia behaved differently in the wild than those without.

In 1989, Koester examined the search and rescue mission reports of two dozen people with dementia. “I sometimes jest that my entire SAR career was originally based upon my ability to average 24 numbers together,” he said.

Despite the study’s modest beginnings, Koester has gone on to larger endeavors—becoming, essentially, the father of a field that he began pursuing in his off hours. In the 2000s, as part of a project he undertook with another researcher that was funded by the U.S. Department of Agriculture, Koester gathered the details of more than 50,000 lost person cases and compiled them into the International Search & Rescue Incident Database, or ISRID, centralizing case details that had previously been scattered.

“I get this panicky phone call from the search planner who’s like, ‘Bob, I absolutely have to know the length of your thumb.”

“He’s just the king of search and rescue,” said Krystal Dacey, a geospatial researcher at Charles Sturt University in Australia. She and other researchers rely on ISRID’s information and Koester’s previous analysis for their own work. “In the SAR research community, it’s mandatory to have as a source,” wrote John Nguyen in an email to Undark. As a student at Virginia Tech, he worked on a software project called the Lost Person Simulation Tool using the database and Koester’s research.

With all that data, Koester began renovating Syrotuck’s work, throwing into relief 41 different lost-person categories. Those human categories range from anglers to ATV drivers, hikers to mountain bikers. “I am sure for some future search there will be a need for statistics regarding female elderly light-wind board-sailers,” he wrote in the 2008 book that resulted from his work, called “Lost Person Behavior.”

For each category, Koester calculated, in quartiles, distances people typically traveled from the point they were last seen, and the elevation change they usually made. He also noted, among other things, how long they remained mobile, how close to features like roads or streams they were found, and what scenario led to their predicament.

School-age kids, for instance, often got lost because they tried to take a shortcut, and can usually be found within a mile of their last known position; more than half are found in structures, yards, or vehicles. Hunters, meanwhile, travel a similar distance, tending to go around 100 feet downhill and to take that journey off-trail—and usually get lost because they followed an animal intently through dense brush. Cross-country skiers, meanwhile, often keep moving to stay warm, and go more than twice as far as hunters or kids—though, some are reported lost when they’ve already made it back to the local bar and failed to tell anyone.

Importantly, Koester also calculated how long people typically live, and how incident coordinators could use the information in real missions.

Search theory, though, is complicated and nuanced. In the field, search and rescue teams can take different approaches when scouting for lost people, but one traditional method is to establish the relevant area by taking the person’s last known position and bounding how far they likely traveled.

Teams may send members out to quickly check features like trails or other areas where a person’s profile suggests they are likely to be found. They search the immediate area, search the boundary of the relevant region, and suss out high-probability spots, like the locations of known hazards. If the lost person isn’t found, teams can fill in the rest of the broad area they’ve established, for example splitting the bounded area into sections that smaller teams can search. Searches also include external information personalized to the situation and terrain.

Koester’s findings also get integrated into courses on best search tactics. The search management curriculum from the Colorado Search and Rescue Association, for instance, focuses on Koester’s “Lost Person Behavior” and the big database. And they’re in the process of revamping it to include new strategies, according to Daniel Knudsen, who teaches search manager classes for the state’s search and rescue organization and is the field director of Park County Search and Rescue in Colorado. Through funding from the Department of Homeland Security, the book’s information now comes in the form of an app, also called “Lost Person Behavior.”

The search for Read, the engineering student in Montana, was managed by the National Park Service, which Koester has worked with. The local SAR teams who helped comb the mountains were involved in “manpower, not decision-making,” in that instance—meaning their members were pounding the ground, but not determining the strategy of their search. But according to Flathead County Search and Rescue president Anthony Palmiotti, “both Flathead and North Valley are very familiar with ‘Lost Person Behavior,’ by Robert Koester, and apply the principles and data in that publication in any search we may manage.” In fact, both teams and the air support had, in December 2022, recently attended a three-day Emergency Response International course together on SAR management and lost person behavior.

Read was lucky: He was found alive.

His journey was chronicled in a Flathead Beacon article : The piece reported that, three days after he went missing, a helicopter team was flying in the area he was last seen, demonstrating how an infrared camera could aid a search. (The flight itself was not part of the official search mission.) And then, they spotted something: Scoot marks in the downslope snow, followed by footprints.

After refueling and returning, the helicopter—with help from search crews—was able to follow those signs and eventually find Read himself, a white-hot spot on the infrared camera in a forest of cool trees. Soon, a rescuer was lowered down to Read, and then Read was lifted into the sky—almost hypothermic, very frostbitten, but alive.

Searchers, and the public, later learned his story: When Read encountered a large, deep field of snow between two high points, he attempted to cross it, but slipped into snow that reached his chest. He’d lost his phone, his water, his shoes.

Though Koester’s methodology has steered the field of search and rescue, there are few independent studies analyzing its effectiveness, and some of the categories in the book have small numbers on which to base their statistics. Still, the ubiquity of its use in real life-or-death situations, formalized SAR education, and further research speaks to the trust both the communities of SAR practitioners and SAR researchers have in Koester’s work.

A National Science Foundation-funded study by Koester, as yet unpublished and not peer-reviewed, estimates that using the lost person behavior strategy has cut search times by around 50 percent. However, research like a 2014 paper led by a University of California, Merced, engineer points to gaps in the idea: The strategies are derived from, essentially, averaging data from across the world, including many places where the geography differs wildly. The paper suggests that in places with “limited geographic diversity,” like a landscape dominated only by rugged mountains, a generalized, smoothed-out strategy might not work as well. The paper specifically looked at Yosemite National Park, part of the Sierra Nevada mountain range and dominated by cliffs and canyons. A 2012 thesis had found something similar for Yosemite, and suggested instead looking at how many watersheds people tended to cross.

In a 2015 publication, researchers from Penn State, George Mason University, Kingston University, and a private consulting firm developed a metric, called MapScore, for evaluating how well a lost-person model performed—in terms of how well it would have predicted a real person’s behavior. In their initial publication , on which Koester is a co-author, they found that Koester’s model outperformed the 2012 watershed model, but combining the two into one idea performed better than both.

Matthew Read built another shelter, near flowing water, and simply waited — an “excellent” strategy, according to “Lost Person Behavior.”

Search and rescue techniques have been getting more refined as people, including Koester, have built on his work in the 16 years since his first book came out. In 2018, Koester co-authored a paper with geographers from the University of Graz in Austria, looking at a model that analyzes features like trails and streams, elevation changes, and behavioral profiles from ISRID to make a probability map displaying different places a lost person might be. He’s also been working with graduate students at University of Virginia on software that takes weather, landscape, and skill level into account when calculating how long people can survive. “Now, it’s just like a matter of finding funding to turn that into an app,” Koester said.

Software has become popular in SAR study. One variety is called an “agent-based model,” in which the “agent” is a digital version of a lost person that interacts with a simulated environment. Koester collaborated with Amanda Hashimoto and others at Virginia Tech on such a model in a 2022 study : It accounts for different “reorientation strategies”—ways people try to get themselves un-lost—which the agents use on their simulated trek through the wilderness. The researchers then compared real ISRID incidents to the paths of the simulated agents. The behavior of the simulated people, on average, fit more than half the real-world incidents well, according to the paper.

A follow-up study , in which Nguyen collaborated with Koester on the Lost Person Simulation tool, assumed that a person tends to keep orienting themselves with the same strategy for a while; it performed even better than Hashimoto and Koester’s model.

Despite the spate of research and new models, though, “there is currently no evidence to suggest that any of the spatial models,” except the one that involves drawing a math-informed circle around someone’s last known position, “have been used in real-time SAR incidents,” wrote Dacey and colleagues in an Australian Journal of Emergency Management study.

As part of her geospatial graduate work, Dacey has been building an agent-based model for practical field use. The “agent” is given rules to abide by—based on the nature of the environment, lost-person behavioral profiles, and a sprinkle of randomness, like an arbitrarily selected starting point and initial direction.

Dacey’s model also incorporates two geographic layers of the search—steepness and the density of vegetation—to approximate ruggedness, and uses them to estimate how long it would take someone to cross a given section. “When you’re lost, and you don’t know where you want to go, the terrain is what’s really driving you,” she said. Koester’s qualitative descriptions of people’s behavior does account for terrain, though the traditional focus in mission planning can tend more toward straightforward numerical variables like elevation change, distance traveled, and the angle walked from the initial starting point.

When the simulation runs, the agent can decide, at each increment of time, whether to move, and in which direction. It keeps track not just of where the agent is and has been, but how long they have been lost, how far they have gone, how tired they are—almost like a video game character whose energy bar is slowly depleting.

Right now, running the model, which could show both where a lost person might be, and a SAR team’s path of least resistance to them, is too computationally intensive for the backcountry, but Dacey is building it with future real-world use in mind; Australian SAR teams still learn to make their search plans—informed by lost person behavior—on paper maps.

One big thing holding Australian SAR teams back from new technology, Dacey says, is that what they have works: They successfully find most lost people. “When there is a time-sensitive thing like a search on, they probably aren’t super confident to go, ‘Okay, well, let’s use this new technology.’ They’re like, ‘Let’s do what we’ve always done.’”

For that reason, Dacey is hoping to prove her model’s worth by using it in cold cases—finding the remains of those people who weren’t found initially, based on reconstructing their behavior.

“I’ve been on searches, and you see the family there,” Dacey said. “And you see how devastated they are.” She’s motivated, she said, to do “anything to help those people.”

Over in the US., Matt Jacobs, a longtime member of multiple California SAR teams, is also trying to bushwhack the way to the future of search theory. Jacobs is an MIT-educated software developer who now runs a company that distributes an app called SARTopo, a mapping software aimed at search and rescue teams.

Using SARTopo, teams can overlay all sorts of geologic and geographic data, plan and plot search areas, and live-track themselves and fellow members, among other functions. In 2015, Jacobs published a paper that took another look at the incident information in the large ISRID database. Taking the largest ISRID categories—hikers, hunters, and gatherers—he tried see how the terrain affected their choices.

Historically, the plan to look for a lost person has sometimes involved a kind of grid search, with evenly spaced people walking in relatively straight lines. That idea calls back to maritime search philosophy, where SAR began: It assumes that searchers can uniformly survey a space where they’re likely to find someone. That may be true on an ocean, but on land, there are cliff bands and dense pockets of deadfall and briar patches that limit where someone may go, and how well they could be spotted. Koester has not advocated for grid searching in his own work.

“When you’re lost, and you don’t know where you want to go, the terrain is what’s really driving you.”

Starting with the locations where people in ISRID were found—and looking at them in isolation, independent of where they were last seen—Jacobs drew an approximately 2-kilometer radius around those spots. Within the resulting circle, he analyzed terrain features near which the lost had been found: things like the intersections of streams and trails, or peaks.

The results were striking, with 60 to 80 percent of people discovered close to terrain features like high and low points, ridges, roads, trails, or streams, as Read was. Injured people, for instance, are 12 times more likely to be found near the interface of a stream and a trail, seven times more likely to be found on a trail, and 3.5 times more likely near just a stream, as compared to a random spot within the circle. Among the uninjured, stream-trail junctions, trails, and roads top the list, with seven times, five times, and three times probability of containing the lost person, respectively.

Given those numerical wins, Jacobs has suggested the idea that broad, grid-like searches over a given area should only happen once SAR teams have picked over relevant geographic features like the intersections of streams and trails. People like Knudsen and Ben Wilson, education director for the Colorado Search and Rescue Association, also now teach this. As the mission ventures farther from the person’s last known position, Jacobs wrote in his 2015 paper, searchers should focus even more on those linear features.

For Koester’s part, he sees Jacobs’ analysis as consistent with the behavioral trends he found in the hiker category and points out that he discusses the importance of terrain features. While Jacobs may put different emphasis or interpretation on how those features play into a search plan, Koester wrote in an email, “I have emphasized linear features as long as I have been doing research. Perhaps it was a bit more obscure in the book than in one of my in-person classes.”

Changing the strategy to one more focused on planetary characteristics doesn’t mean abandoning lost person behavior, said Michael St. John, who leads the Search and Rescue Unit at the Marin County Sheriff’s Office in California. Together with Jacobs, St. John developed and teaches a course, certified by the National Association for Search and Rescue, that incorporates Jacobs’s research and the SARTopo software itself. This allows the whole team to GPS-track themselves on one map that search managers can also see—and so keep track of where the team is, determine well an area is covered, and give direction from their big-picture position in the office, as long as the team members have cell service.

Within the rings, he and Jacobs suggest in a new version of the “lost person incident management” course they teach, teams should focus on the most promising terrain features, rather than focusing on searching a delineated area. “This is now, I think, a more holistic approach,” agreed Wilson.

GPS tracking of team members is one tool; telling them where to look based on the terrain is another; lost person behavioral profiles, St. John says, are another. SAR, complicated as the Wild West, needs them all to account for the strange and dangerous situations in which people find themselves. “There’s going to always be outliers,” said St. John.

Soon, search efforts will have more data and analysis, to sharpen those tools: Koester has been gathering new case reports, ISRID now has around 300,000 incidents, and “Lost Person Behavior” is getting a second edition. Koester guesses the number of person-categories will shoot up to around 75, including new ones like BASE jumpers—people who leap from objects like bridges, buildings, or cliffs—that are wearing wingsuits.

It’s not quite female elderly light-wind board sailors, but it’s homing in on that destination.

Like science, tech, and DIY projects?

Sign up to receive Popular Science's emails and get the highlights.

We use research to design groundbreaking strategies that authentically reach high-risk audiences.

IN-HOUSE BEHAVIOR CHANGE RESEARCH

Rescue’s 20+ member in-house research team serves as the driving force for our behavior change interventions. Composed of MPH and PhD research scientists, our team leads formative, segmentation, and message testing research for all Rescue campaigns and interventions. By keeping research in-house, we ensure that research findings are woven into every aspect of each intervention, from key messages to implementation tactics.

Rescue’s research approach is built based on nearly two decades of effective behavior change interventions. This experience allows us to design research studies that provide practical and actionable insights for our targeted populations.

PSYCHOGRAPHIC SEGMENTATION

PEER CROWDS

CREATIVE AND MESSAGE TESTING

PUBLICATIONS

Why Peer Crowds Matter: Incorporating Youth Subcultures and Values in Health Education Campaigns American Journal of Public Health

Peer Crowd Identification and Adolescent Health Behaviors: Results From a Statewide Representative Study Health Education & Behavior

Freedom From Chew: Using Social Branding to Reduce Chewing Tobacco Use Among Country Peer Crowd Teens Health Education & Behavior

To learn more about our strategies, please contact:

Learn More About Our Behavior Change Strategies:

Learn more through our case studies.

A youth engagement initiative to prevent tobacco use on school grounds

A health communications campaign working to educate Vermont young adults about the dangers of recreational opioid use

A tobacco prevention campaign targeting high-risk teens to associate Country values with a tobacco-free lifestyle

• Case Studies
• Commercials
• Press & Awards
• Decision Blocks
• Social Branding
• SAVI Messaging

An official website of the United States government

The .gov means it’s official. Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

The site is secure. The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

• Publications
• Account settings

Preview improvements coming to the PMC website in October 2024. Learn More or Try it out now .

• Advanced Search
• Journal List
• Elsevier - PMC COVID-19 Collection

The effect of the rescue plans and the need for policies to increase economic growth ☆

This paper evaluates the economic impact of discretionary fiscal and monetary actions taken in the United States during 2020 and 2021. The fiscal actions are The Coronavirus Aid, Relief, and Economic Security Act, or the “CARES” Act, passed in March 2020; The Consolidated Appropriations Act, passed in December 2020; and The American Rescue Plan Act, passed in March 2021. The paper focuses on the impact of the “economic impact payments” that underlie these fiscal actions. The paper also examines discretionary monetary policy actions taken during the same period. The overall implication is that there is a need to return to policies that increase economic growth and stability, including rules-based fiscal and monetary policy, rather than to continue with these one-time discretionary actions.

1. Introduction

During the period from March 2020 to March 2021 the United States government enacted three fiscal packages with a main purpose of stimulating the American economy during and after the pandemic that first hit the world in 2020. In each case, an amount called the Economic Impact Payments (EIP) was distributed to people in the United States by various methods, including direct deposit, checks, or special prepaid debit cards. The term “Economic Impact Payments” was introduced in 2020 for what had previously been called “stimulus checks” or “recovery rebates” as explained by the Consumer Financial Protection Bureau (2021) . The idea was that people would spend the Economic Impact Payments, increase their personal consumption expenditures and thereby increase aggregate demand and stimulate the economy. The stated purpose of the Economic Impact Payments was to stimulate the economy though some assistance in the overall fiscal packages was justified on humanitarian grounds.

There is no question that the Economic Impact Payments increased disposable personal income, as immediately reported in news outlets (see Cambon, 2021 ). The correlation between disposable personal income and the Economic Impact Payments was very high at.86 from January 2020 to November 2021. The increase in the Economic Impact Payments in April 2020, January 2021 and March 2021 was virtually all of the increase in disposable personal income.

But it is less clear that these payments increased consumption—as had been argued in advance of the legislation and continues to be argued today. Such an increase in consumption was needed, according to many macroeconomic theories, to stimulate the economy. Here I examine whether the fiscal actions actually impacted consumption. I also examine the move to discretionary monetary policy following the pandemic, and raise questions about the impact of these policies on inflation and economic growth.

The overall implication of the paper is that we need to return to rules-based policies that increase economic growth and stability rather than continue with these one-time discretionary actions. The empirical method is similar to that used by Cogan, Cwik, Taylor, and Wieland (2010) , Gramlich (1979) , and Taylor (2009) , and to examine fiscal packages and to that used by Cochrane, Taylor, and Wieland (2020) to evaluate monetary policy rules.

2. The economic impact of discretionary Fiscal Policy in 2020 and 2021

The first of the three recent fiscal packages, called the Coronavirus Aid, Relief and Economic Security Act of 2020 or simply the CARES Act, was passed by the United States Congress and signed into law by President Donald Trump on March 27, 2020. It provided direct support through Economic Impact Payments to individuals, including advance tax rebate payments distributed in April 2020. The second package, called The Coronavirus Response and Relief Supplemental Appropriations Act of 2021 was passed and signed into law by President Donald Trump on December 27, 2020. It provided a second round of Economic Impact Payments to individuals. The third package, called The American Rescue Plan Act of 2021 or the COVID-19 Stimulus Package, was passed by Congress and signed into law by President Joe Biden on March 11, 2021. It provided a third round of Economic Impact Payments.

Economic Impact Payments are a specific amount of dollars ranging, over the three fiscal packages, from $600 to $1400 for an individual, from $1200 to $2800 for married taxpayers jointly filing, and from $500 to $1400 additional for dependents. In each of the three pieces of legislation, the Economic Impact Payments had an income threshold above which the payment was reduced, and it was eventually phased out completely for people with incomes above $150,000 or $160,000 for married taxpayers filing jointly. For those who filed tax returns, the payments were made automatically just as tax refunds. The majority of payments were made through direct deposit, check, or through a pre-paid debit card issued by the Department of Treasury. The card was sent in a white envelope from the “Economic Impact Payment Card” with the U.S. Department of the Treasury seal displayed.

A summary of the Economic Impact Payments over the months of 2020 and 2021 is shown graphically in Fig. 1 using calculations made by the Bureau of Economic Analysis of the Department of Commerce. These monthly data on the Economic Impact Payments are reported in various monthly publications of “Effects of Selected Federal Pandemic Response Programs on Personal Income.” The latest publication was dated December 23, 2021, and it reported data through November 2021. The three packages are labeled in Fig. 1 by the names of the legislation. The majority of payments for the third act were distributed in March 2021, though the Internal Revenue Service made some smaller additional payments later in the year as tax returns were processed.

Economic Impact Payments (EIP) in Three Fiscal Packages: 2020–2021, (Billions of Dollars, Seasonally Adjusted Annual Rates, Monthly Data, From “Effect of Selected Federal Pandemic Response Programs, various issues).

In each case the Economic Impact Payments were meant to be temporary for the period shown in Fig. 1 . It was assumed that each of the Economic Impact Payments would be phased out as shown in the graph. Moreover, there was no notice given in March 2020 that there would be more payments in January 2021 or March 2021.

The rationale for such temporary payments is that they increase disposable personal income and thereby increase the demand for consumption, aggregate demand and the whole economy. This rationale is based on the Keynesian consumption function in which an increase in income, increases consumption, which adds to total demand and increases GDP. An alternative view, the permanent income theory of Friedman (1957) or life-cycle theory of consumption of Modigliani (1976) , stresses that such temporary increases in income lead to very small increases in consumption in comparison with more permanent increases.

Fig. 2 shows the impact of the Economic Impact Payments on disposable personal income. The upper line shows disposable personal income for the months from January 2017 through November 2021. The data are seasonally adjusted and are stated at annual rates. Disposable personal income is the total amount of income after taxes and government transfers, and it therefore includes the Economic Impact Payments. Subtracting the Economic Impact Payments from personal consumption expenditures shown in the upper line results in the lower line in Fig. 2 labeled “Without Economic Impact Payments.” It shows what disposable personal income would have been without the Economic Impact Payments. Notice the sharp increase in disposable personal income when the Economic Impact Payments were made. Disposable personal income then comes down sharply as the Economic Impact Payments decline, and the amount then returns to its original level. Prior to the Economic Impact Payments, disposable personal income grew at a steady pace as shown in Fig. 2 .

Disposable personal income with and without Economic Impact Payments.

Fig. 3 shows the pattern of personal consumption expenditures along with disposable personal income and the Economic Impact Payments. Note that consumption shows no increase at the time of the Economic Impact Payments in March 2020. Consumption fell sharply as major parts of the economy were shut down due to shelter-in-place restrictions and concern about the spread of the coronavirus. Similarly, there was little or no impact on consumption in January 2021 or March 2021 when there were huge increases in disposable personal income due to the Economic Impact Payments. Moreover, it is not plausible to say that consumption would have declined had these payments not been made, because the economy was not in a recession and there was nothing separately pulling consumption down in either January 2021 or March 2021. As Fig. 3 shows that the temporary payments did little to stimulate consumption.

Economic impact payments and consumption.

We can also look at personal saving, defined as disposable personal income less personal consumption expenditures, interest payments, and current transfer payments to government and the rest of the world. The personal saving rate is shown in Fig. 4 . The saving rate rises sharply and temporarily each time that Economic Impact Payments are made. It then falls back again after the payment is over. The chart clearly demonstrates that a large amount of the Economic Impact Payments was actually saved rather than spent on consumption, as one would predict from the permanent income or life-cycle theories.

Personal Saving Rate: January 2017 to November 2021.

While the graphs in Figs. 3 and ​ and4 4 are clear about the small impacts of the Economic Impact Payments on consumption, testing for the impact on aggregate consumption using statistical techniques provides clearer evidence, as shown by regression estimates in Table 1 , in which personal consumption expenditures is the dependent variable. The regression covers January 2017 through November 2021, and includes all payments as in Fig. 1 , Fig. 2 , Fig. 3 .

Personal Consumption Expenditures (PCE), Economic Impact Payments (EIP) and Disposable Personal Income (DPI).

R 2 = 0.88.

To test whether the Economic Impact Payments have a significant effect on consumption, the regression includes: (1) personal disposable income without the Economic Impact Payments and (2) the actual Economic Impact Payments. To allow for partial adjustment of consumption to the changes in income, a lagged dependent variable is included.

As one can see from Fig. 3 , the actual Economic Impact Payments are far more volatile than the personal disposable income without the payments. According to permanent income or life-cycle theories, the effect on consumption of these temporary effects should be small.

Table 1 shows that the consumption impact of the Economic Impact Payments is statistically insignificant. In contrast, the coefficient on disposable personal income excluding the rebate is significant. This confirms the results illustrated in Fig. 3 . In these regressions a temporary increase in income—represented by the Economic Impact Payments variable—has a small and statistically insignificant effect. When the increase in income is more permanent—as represented here by the personal disposable income variable without the Economic Impact Payment— then the change in consumption is larger and statistically significant. The coefficient on the Economic Impact Payments, though statistically insignificant, is negative, which may reflect a missing third variable. It could be that some forces, perhaps indirectly related to the pandemic, caused the decline in consumption and also an increase in payments.

The recent experiences with the Economic Impact Payments is not the first time that the government has endeavored to stimulate the economy with temporary increases in income, and it is very informative to compare the difference over time. Fig. 5 shows the impact of the one-time rebate in 2008. The upper line shows monthly disposable personal income, seasonally adjusted at an annual rate. Disposable personal income includes the rebate payments in 2008 just as personal disposable income includes the Economic Impact Payments in 2020 and 2021. Subtracting the rebate payments from disposable personal income results in disposable personal income without the rebate, also shown in Fig. 5 .

Income with and without Rebate and Consumption: 2007–08.

Notice that there is a sharp increase in disposable personal income when rebates were distributed. Disposable personal income then declines as the rebate declines, and eventually returns to the previous trend. Personal consumption expenditures shows no increase at the time of the rebate. The temporary rebate did little or nothing to stimulate consumption demand, and thereby aggregate demand, or the economy. The graph is much the same as in 2020 and 2021.

While the data shown in Fig. 5 tell a clear story, we can also look at the impact of the rebates payments on aggregate consumption using regression techniques as in Table 1 . Here it is useful to include also a rebate payment made in 2001, which is shown along with the 2008 rebate payment in Fig. 6 . Both rebates were temporary and the effect on consumption should be relatively small. We thus start the sample period in January 2000 and go through October 2008, and thereby include both the rebate payment in 2008 and a similar rebate payment in 2001.

Rebate Payments in 2001 and 2008. Source: Monthly data from BEA Publication “Personal Income and Outlays,” September 2001, October 2001, August 2008.

The results are shown in Table 2 . To test whether the rebates had a positive and significant effect on consumption, both personal disposable income without the rebates and the rebate payments are two separate variables in the regressions. Table 2 shows that the impact of the rebate is statistically insignificant while there is a significant impact of the more permanent disposable personal income excluding the rebate. The results are remarkably similar to Table 1 .

Personal Consumption Expenditures (PCE), Rebates and Disposable Personal Income (DPI).

R 2 = 0.99.

The earlier packages differed in size, duration, and the mechanism for distribution of the payments compared with the 2020 and 2021 packages, but they were quite similar from a macroeconomic perspective. They were widely viewed as temporary and were justified mostly on the grounds of jump-starting consumption and stimulating the economy. In fact, a major principle underlying the 2001 and 2008 stimulus packages was that they should be temporary, as well as targeted and timely. This temporary feature distinguishes these actions from more permanent changes such as the personal income tax rate cuts in the 1960s and 1980s

3. The economic impact of discretionary monetary policy

The pandemic that started in the first quarter of 2020 with COVID-19 was a jolt to the economy. It interrupted a revival of rules-based policies as the Fed took special actions to deal with the effects of a health crisis on the economy. These actions included a rapid reduction in the target for the federal funds rate from 1.75 % to 0.25 % during the weeks of March 2020. Also included were large-scale purchases by the Fed of Treasury and mortgage backed securities causing a large expansion of the Fed’s balance sheet, from $3.8 trillion before the pandemic to $8.8 trillion. Both M1 and M2 measures of the money supply grew rapidly. These actions were discretionary and were not consistent with rules-based policies. The Fed also stopped reporting on rules-based policy in its Monetary Policy Report in the July 2020.

However, the Federal Reserve’s departure from reporting on rules-based policy was short-lived. The Monetary Policy Reports that were released on February 19, 2021 and July 9, 2021, again included a section on monetary policy rules. That policy rules reentered the Report was a welcome development. It re-initiated helpful reporting that began in the July 2017 Monetary Policy Report, but was dropped in July 2020.

Five rules were discussed in the February and July 2021 Monetary Policy Reports. To quote the Reports, these include “the well-known Taylor (1993) rule, the ‘balanced approach’ rule, the ‘adjusted Taylor (1993)’ rule, and the ‘first difference’ rule. In addition to these rules,” and this is very important, there is a new “‘balanced approach (shortfalls) rule,’ which represents one simple way to illustrate the Committee’s focus on shortfalls from maximum employment.”

Table 3 shows the five rules from the 2021 Reports: There were also five rules in the earlier Reports, but one was left out, and a new one—the Balanced-approach (shortfalls) rule—was added. As stated in the document, this modified simple rule “would not call for increasing the policy rate as employment moves higher and unemployment drops below its estimated longer-run level.”

Five policy rules in the February 2021 monetary policy report.

Note: The equations represent the values of the nominal federal funds rate prescribed by the Taylor (1993), balanced approach, balanced-approach (shortfalls), adjusted Taylor (1993), and first difference rules, respectively.

It is good that rules were put back in the Fed’s Monetary Policy Report, but it would be much more helpful if the Fed also incorporated rules into its actual decisions. Apparently this did not happen, as a comparison of the interest rate path and policy rules for the interest rate in Fig. 7 suggests. Fig. 7 compares an FOMC projection of the federal funds rate out to 2023Q4 and three different rules-based paths for the federal funds rate to 2023.

FOMC projection of federal funds and monetary policy rules with three inflation rates.

The three lines in Fig. 7 show the federal funds rates from three policy rules using the same parameters as those in the Taylor rule, which is discussed in the February 2021 Monetary Policy Report. The so-called equilibrium interest rate has been reduced from 2 % to 1 % in the calculations in Fig. 7 . Such a reduction has been suggested at the Fed. The three policy rules use the four-quarter inflation rates of the GDP price index, the PCE price index, or the core PCE price index, based on the February 2021 Congressional Budget Office (CBO) projections. They use the same percentage deviation of real GDP and from potential GDP as the CBO report.

Even with this smaller equilibrium real interest rate (1 % rather than 2 % in the original Taylor rule), the FOMC’s path for the federal funds rate is well below any of these policy rules. There was already a difference in the first quarter of 2021, and the difference grew over time. Consider for simplicity the average of the interest rates for the three different inflation rates in the final quarter of each year. If we average the three values, we get 1.9 % in 2021Q4, 2.5 % in 2022Q4 and 2.7 % in 2023Q4.

There was little mention of why the discrepancy existed between the Fed’s actual decisions reported here and the rules. Does this mean that the Fed will keep the rate this low under similar circumstances regarding real GDP and inflation? The rate is exceptionally low compared to similar periods in recent history. Higher possible levels for the federal funds rate were not mentioned in the Fed’s reported discussions during this period. Members of the FOMC argued that the higher inflation was transitory due to pandemic’s effect which brought inflation down in 2020.

Those who defended this stance pointed out that market interest rates on longer-term bonds remained very low. On safe Treasury assets, in August 2021, the five-year yield was only 0.81 %, and the ten-year yield was only 1.35 % – well below the rates suggested by rules such as the Taylor rule when averaged over these maturities. Considering these factors, people argued that the markets were being rational when they forecast low rates.

The problem with this line of reasoning is that the low longer-term rates were likely being caused by the Fed’s own insistence on keeping low rates. Effectively, the policy rule for longer-maturity bonds depends on the policy rule for the much shorter-term federal funds rate, as perceived by people in the market. If the Fed convinces the market that it will stay low, the term structure of interest rates will imply lower longer-term rates.

The situation was similar to that of 2004, when then-Fed Chair Alan Greenspan noticed that ten-year Treasury yields did not seem connected to moves in the federal funds rate.

But one need look no further than the Fed’s own July 9, 2021, Monetary Policy Report, which includes long-studied policy rules that would prescribe a policy rate higher than the current actual rate. As inflation picked up, the Fed argued that the high inflation simply reflected the bounce back from the low inflation of last year.

4. Conclusion: the need for rule-like policies to increase economic growth

This paper has examined the reasons to return to a rules-based fiscal and monetary policy in the United States and has outlined a method to do so. By reviewing the ineffectiveness of deficit-financed discretionary fiscal policy and the inflation-inducing effects of a discretionary monetary policy in the past few years, the paper provides the background needed for analyzing current and future fiscal and monetary policy decisions.

The results indicate that policy should now be strategic or rule-like in which people and markets understand that the fiscal authorities aim for a steadiness in policy without large discretionary deficit-expanding actions and the monetary authorities raise the policy interest rate as inflation rises. It would be a contingency plan, as all rules and strategies should be. By having clearly discussed policy rules in its February 2021 and July 2021 Monetary Policy Reports, the Fed has prepared for such a strategy in practice. Explaining how its policy rule or strategy would be consistent with its flexible average inflation targeting statements would further clarify the Fed’s monetary policy and facilitate the market adjustment when it takes place. It would remove uncertainty and remaining inconsistencies.

These implications are the logical conclusion of many years of macroeconomic experience. Two decades ago, Eichenbaum (1997) wrote, “There is now widespread agreement that countercyclical discretionary fiscal policy is neither desirable nor politically feasible.” Feldstein (2002) likewise argued “There is now widespread agreement in the economics that deliberate ‘countercyclical’ discretionary policy has not contributed to economic stability and may have actually been destabilizing in the past.” And in Taylor (2000) , in a paper, entitled “Reassessing Discretionary Fiscal Policy,” I concluded similarly that “in the current context of the US economy, it seems best to let fiscal policy have its main countercyclical impact through the automatic stabilizers…”.

Despite this widespread agreement, there was a dramatic revival of interest in discretionary fiscal policy in 2008, including temporary rebates and stimulus checks designed to stimulate the economy. I examined these policies in detail and found in Taylor (2009) and then summarized in Taylor (2011) that “once again these temporary discretionary one-time countercyclical payments did nothing to stimulate the economy” and “raise considerable doubts about the efficacy of temporary discretionary countercyclical fiscal policy in practice’ which “adds more weight to the position reached… by Gramlich, 1978 , Gramlich, 1979 and Lucas and Sargent (1978) .” Cogan, Taylor, and Wieland (2009) published an article in the Wall Street Journal summarizing the evidence and concluding, as the article’s title said, that “The Stimulus Didn't Work.” As a result, if this research and that of others, monetary and fiscal policy began to turn toward a rule like approach in the years from 2017 to 2019, but the pandemic that began in 2020 brought back another wave of discretionary fiscal and monetary policy.

The purpose of this paper has been to review carefully—using econometrics, simple charts and data—the empirical evidence on this return to discretion fiscal and monetary policy in the years 2020 and 2021. The review confirms the view put forth two decades ago, and one decade ago, and at many other times and places in the United States and elsewhere that these temporary programs—whether fiscal or monetary—do not stimulate the overall economy and have their own negative effects.

☆ George P. Shultz Senior Fellow in Economics at the Hoover Institution, Stanford University, and Mary and Robert Raymond Professor of Economics, Stanford University. This paper was prepared for the session: “Will U.S. Growth Be Higher than in the Previous Decade after the Pandemic Fiscal Stimulus Ends?” at the American Economic Association on January 7, 2022, chaired by Dominick Salvatore. The paper uses data from the Bureau of Economics Analyis (BEA), though, as stated by th BEA in the December 23, 2021 release, some of “the impacts are generally embedded in source data and cannot be separately identified.” https://www.bea.gov/sites/default/files/2021-12/pi1121.pdf .

Appendix. : Description of the three rounds of Economic Impact Payments (EIP)

Source: Consumer Financial Protection Bureau, CFPB (2021) “Your Guide to Economic Impact Payments,” in section called EIP round differences based on your income and dependents.

Third round of EIPs (issued starting in March 2021) .

PAYMENT AMOUNT.

$1400 per eligible individual.

$2800 for married joint filers.

Additional $1400 for each dependent of all ages, including children under 19, college-age students, and adults with disabilities.

INCOME THR„E„S„H„OLD.

Total payment phases out between following income levels:

$75,000-$80,000 singles.

$112,500-$120,000 heads of households.

$150,000-$160,000 married filing jointly.

Second round of EIPs (issued starting in December 2020) .

$600 per eligible individual.

$1200 for married joint filers.

Additional $600 for each qualifying child or dependent child under 17 years old.

Total payment amount will be phased out by $5 per $100 of income above these thresholds:

$75,000 singles.

$112,500 heads of households.

$150,000 married filing jointly.

First round of EIPs (issued starting in March 2020) .

$1200 per eligible individual.

$2400 for married joint filers.

Additional $500 for each qualifying child or dependent under 17 years old.

• Cambon S.C. Spurred by checks, household income rises a record 21.1 % Wall Street Journal. 2021 p. A2. [ Google Scholar ]
• CFPB (2021). Your guide to economic impact payments . 〈https://www.consumerfinance.gov/coronavirus/managing-your-finances/guide-economic-impact-payments/〉 .
• Cochrane J.H., Taylor J.B., Wieland V. In: Strategies for monetary policy. Cochrane J., Taylor J., editors. Hoover Institution Press; 2020. Evaluating rules in the fed's report and measuring discretion; p. 2020. [ Google Scholar ]
• Cogan J.F., Cwik T., Taylor J.B., Wieland V. New Keynesian versus old Keynesian government spending multipliers. Journal of Economic Dynamics and Control. 2010; 34 (3):281–2950. March 2010. [ Google Scholar ]
• Cogan J.F., Taylor J.B., Wieland V. The stimulus didn't work. Wall Street Journal. 2009; 17 :2009. 〈https://web.stanford.edu/~johntayl/2009_pdfs/The-Stimulus-Did-not-Work-WSJ-09-17-09.pdf〉 [ Google Scholar ]
• Eichenbaum M. Some thoughts on practical stabilization policy. American Economic Review. 1997; 87 (2):236–239. [ Google Scholar ]
• Feldstein, M. (2002), T he role for discretionary fiscal policy in a low interest rate environment, National Bureau of economic research working paper 9203 .
• Friedman M. Princeton University Press; 1957. A theory of the consumption function. [ Google Scholar ]
• Gramlich E.M. State and local budgets the day after it rained: Why is the surplus so high? Brookings Papers on Economic Activity. 1978; 1 :191–214. [ Google Scholar ]
• Gramlich E.M. Stimulating the macro economy through state and local governments. American Economic Review. 1979; 69 (2):180–185. [ Google Scholar ]
• Lucas R.E., Sargent T.J. After the phillips curve: Persistence of high inflation and high unemployment. Federal Reserve Bank of Boston; Boston: 1978. After Keynesian macroeconomics; pp. 49–72. [ Google Scholar ]
• Modigliani F. Life-cycle, individual thrift, and the wealth of nations. American Economic Review. 1976; 76 (3):297–313. [ PubMed ] [ Google Scholar ]
• Taylor J.B. Reassessing discretionary fiscal policy. Journal of Economic Perspectives. 2000; 14 (3):21–36. [ Google Scholar ]
• Taylor J.B. The lack of an empirical rationale for a revival of discretionary fiscal policy. American Economic Review. 2009; 99 (2):550–555. May 2009. [ Google Scholar ]
• Taylor J.B. An empirical analysis of the revival of fiscal activism in the 2000s. Journal of Economic Literature. 2011; 49 (3):686–702. September 2011. [ Google Scholar ]

New Paper Published:

Caspian terns saved, rehabilitated, and released by international bird rescue are surviving and breeding.

The research paper was co-authored by Julie Skoglund, Rebecca Duerr, DVM MPVM PhD, both of International Bird Rescue, and Dr. Charlie Collins, Professor Emeritus at California State University of Long Beach.

The story began in 2006 and 2007 in the Port of Long Beach, one of the busiest shipping ports on the west coast and near a favored breeding colony locale for both Caspian and Elegant Terns in southern California. In both years, disastrous events threatened the lives of tern chicks born in the Port of Long Beach.

In 2006, workers cleaning the deck of a barge deliberately flushed Caspian Tern chicks-too young to survive independently-into the Pacific Ocean. In 2007, suspected human disturbances caused another group of tern chicks to wind up floundering in the water. Fortunately, Bird Rescue was able to rescue some of these young birds and take them into care at its Los Angeles Wildlife Center.

To understand how we solved the challenges of rehabilitating these terns, please read Survival and Recruitment of Rehabilitated Caspian Terns in Southern California .

The final paper was published in the Bulletin of the Southern California Academy of Sciences, May 2020 .

Also read: Full research paper blog post and Rare Tern Colony Decimated in Long Beach, CA

More Bird Rescue Published Research Papers:

Bamac OE, Rogers KH, Arranz-Solís D, Saeij JPJ, Lewis L, Duerr R, Skoglund J, Peronne L, Mete A. 2020.Protozoal encephalitis associated with Sarcocystis calchasi and S. falcatula during an epizootic involving Brandt’s cormorants (Phalacrocorax penicillatus) in coastal Southern California, USA. International Journal for Parasitology: Parasites and Wildlife 12: 185-191. https://doi.org/10.1016/j.ijppaw.2020.06.005

Horgan MD, Siksay SE, Knych HK, Duerr RS. (in press 2020). Pharmacokinetics of a single dose of oral meloxicam in Brown Pelicans (Pelecanus occidentalis). Journal of Avian Medicine and Surgery.

Piatt JF, Parrish JK, Renner HM, Schoen SK, Jones TT, Arimitsu ML, Kuletz KJ, Bodenstein B, García-Reyes M, Duerr RS, Corcoran RM, Kaler RSA, McChesney GJ, Golightly RT, Coletti HA, Suryan RM, Burgess HK, Lindsey J, Lindquist K, Warzybok PM, Jahncke J, Roletto J, Sydeman WJ. 2020. Extreme mortality and reproductive failure of Common Murres resulting from the northeast Pacific marine heatwave of 2014-2016. PLoS ONE 15(1): e0226087. https://doi.org/10.1371/journal.pone.0226087

Skoglund JM, Duerr RS, Collins, C. 2020. Survival and recruitment of rehabilitated Caspian Terns in southern California. Bulletin of the Southern California Academy of Sciences 119(1): 1-5. Available at: https://scholar.oxy.edu/handle/20.500.12711/10544

Duerr RS. 2018. Successful surgical management of pouch and bill injuries in pelicans. Pp. 363-367 in Proceedings of the Association of Avian Veterinarians annual conference at ExoticsCon, Sept. 22-27, 2018, Atlanta, GA.

Gibble C, Duerr R, Bodenstein B, Lindquist K, Lindsey J, Beck J, Henkel L, Roletto J, Harvey J, and R Kudela. 2018. Investigation of a largescale Common Murre (Uria aalge) mortality event in California in 2015. Journal of Wildlife Diseases 54(3):569-574. https://doi.org/10.7589/2017-07-179

Duerr RS and KC Klasing. 2017. Effects of added lipids on digestibility and nitrogen balance in oiled Common Murres (Uria aalge) and Western Grebes (Aechmophorus occidentalis) fed four formulations of a critical care diet. Journal of Avian Medicine and Surgery 31(2): 132-141.

Duerr RS. 2016. Surgical repair of keel lesions and lacerations in aquatic birds. Pp. 83-88 in Proceedings of the Association of Avian Veterinarians annual conference at ExoticsCon, Aug. 27-Sept. 1, 2016, Portland OR.

Duerr RS, Ziccardi MH, and JG Massey. 2016. Investigation into mortalities during treatment: Factors affecting oiled Common Murre (Uria aalge) survival to release through rehabilitation. Journal of Wildlife Diseases 52(3): 495-505. https://doi.org/10.7589/2015-03-054

Duerr RS and KC Klasing. 2015. Tissue component and organ mass changes associated with declines in body mass in three seabird species received for rehabilitation in California. Marine Ornithology. 43(1): 11-18. http://marineornithology.org/PDF/43_1/43_1_11-18.pdf

Gaynor AS, Fish S, Duerr RS, Dela Cruz Jr. FN, and P Pesavento. 2015. Identification of a novel papillomavirus in a Northern Fulmar (Fulmarus glacialis) with viral production in cartilage. Veterinary Pathology 52(3):553-61. https://doi.org/10.1177/0300985814542812

Li L, Pesavento PA, Gaynor AM, Duerr RS, Phan TG, Zhang W, Deng X, and E Delwart. 2015. A gyrovirus infecting a sea bird. Archives of Virology. 160(8): 2105-9. https://dx.doi.org/10.1007/s00705-015-2468-1

Humple D.L. and Holcomb J. 2014. Winter movements of Western Grebes and Clark’s Grebes: Insight from band recoveries. North American Bird Bander 39(1): 21-28. For full text click here

Duerr RS. 2013. Investigation into the Nutritional Condition and Digestive Capabilities of Seabirds during Rehabilitation in California. PhD Dissertation. University of California, Davis. 119 pp.

Ruoppolo V., E.J. Woehler, K Morgan, and C.J. Clumpner. 2012. Wildlife and Oil in the Antarctic: a recipe for cold disaster. Polar Record 48:1-13.

Duerr RS, Massey JG, Ziccardi MH, and YN Addassi. 2011. Physical effects of Prudhoe Bay crude oil water accommodated fractions (WAF) and Corexit 9500 chemically enhanced water accommodated fractions (CEWAF) on Common Murre feathers and California sea otter hair. California Dept. of Fish and Wildlife, Scientific Study and Evaluation Program, Report 2007-01. 12 pp.

Phillips EM, Zamon JE, Nevins HM, Gibble C, Duerr R, and L Kerr. 2011. Summary of birds killed by a harmful algal bloom along south Washington and north Oregon coasts during October 2009. Northwestern Naturalist 92(2):120-126. https://doi.org/10.1898/10-32.1

Nevins H., M. Miller, L. Henkel, D. Jessup, N. Carion, C. Meteyer, K. Schuler, J. St. Leger, L. Woods, Skoglund J. and D. Jaques. 2011. Summary of unusual stranding events affecting Brown Pelican along the US Pacific Coast during two winters, 2008-09 and 2009-10. Final Report, Marine Wildlife Veterinary Care and Research Center, Santa Cruz, CA.30 pp. For full text go here

Phillips E.M., J.E. Zamon, H.M. Nevins, C. Gibble, R. Duerr, and L. Kerr. 2011. Summary of birds killed by a harmful algal bloom along south Washington and north Oregon coasts during October 2009. Northwestern Naturalist 92(2):120-126. For full text go here

Holcomb J. Overview of bird search and rescue, response efforts during the 1989 Exxon Valdez oil spill. International Oil Spill Conference Proceedings: March 1991, Vol. 1991, No. 1, pp. 225-228. For full text go here

Other papers of interest:

Henkel LA, Ziccardi MH. 2018. Life and death: How should we respond to oiled wildlife? Journal of Fish and Wildlife Management 9(1):296-301; e1944-687X. doi:10.3996/062017-JFWM-054. https://fwspubs.org/doi/full/10.3996/062017-JFWM-054

Newman, SH, RT Golightly, EN Craig, HR Carter and C Kreuder. 2004. The effects of petroleum exposure and rehabilitation on post-release survival, behavior and blood indices: A Common Murre (Uria aalge) case study following the Stuyvesant petroleum spill. Final Report. Oiled Wildlife Care Network, Wildlife Health Center, School of Veterinary Medicine, University of California, 1 Shields Ave, Davis CA 95616. 46 pp. Read full report

Computational Theories of Interaction and Agency

Edited by philip e. agre university of california, san diego stanley j. rosenschein teleos research mit press, 1996.

• Share full article

For Afghan Refugees, a Choice Between Community and Opportunity

In resettling thousands of displaced Afghans, the Biden administration must weigh their need for support against the needs of the U.S. labor market.

Fremont, Calif., is home to a large Afghan enclave known as Little Kabul. On weekends, many Afghan families gather for picnics and walks. Credit... Gabriela Bhaskar/The New York Times

Supported by

By Michael D. Shear and Jim Tankersley

• Nov. 24, 2021

FREMONT, Calif. — Harris Mojadedi’s parents fled Afghanistan’s communist revolution four decades ago and arrived as refugees in this San Francisco suburb in 1986, lured by the unlikely presence of a Farsi-speaking doctor and a single Afghan grocery store.

Over the decades, as more refugees settled in Fremont, the eclectic neighborhood became known as Little Kabul, a welcoming place where Mr. Mojadedi’s father, a former judge, and his wife could both secure blue-collar jobs, find an affordable place to live and raise their children surrounded by mosques, halal restaurants and thousands of other Afghans.

“When I went to school, I saw other Afghan kids. I knew about my culture, and I felt a sense of, like, that my community was part of Fremont,” Mr. Mojadedi recalled recently over a game of teka and chapli kebabs during lunch with other young Afghans from the area.

But now, as the United States begins to absorb a new wave of refugees who were frantically evacuated from Kabul in the final, chaotic days of America’s 20-year war in Afghanistan, it is far from clear that a place like Fremont would be an ideal destination for them. Housing in the Bay Area city is out of reach, with one-bedroom apartments going for more than $2,500 a month. Jobs can be tougher to get than in many other parts of the country. The cost of living is driven up by nearby Silicon Valley. Even longtime residents of Little Kabul are leaving for cheaper areas.

The alternative is to send the refugees to places like Fargo, N.D., or Tulsa, Okla., where jobs are plentiful, housing is cheap and mayors are eager for new workers.

But those communities lack the kind of cultural support that Mr. Mojadedi experienced. The displaced Afghans would most likely find language barriers, few social services and perhaps hostility toward foreigners. Already, there are signs of a backlash against refugees in some of the states where economic statistics suggest they are needed most.

“Are we setting them up to fail there?” Homaira Hosseini, a lawyer and Afghan refugee who grew up in Little Kabul, asked during the lunch. “They don’t have support. Or are we setting them up to fail in places where there aren’t any jobs for them, but there is support?”

That is the difficult question facing President Biden’s administration and the nation’s nonprofit resettlement organizations as they work to find places to live for the newly displaced Afghans. As of Nov. 19, more than 22,500 have been settled, including 3,500 in one week in October, and 42,500 more remain in temporary housing on eight military bases around the country, waiting for their new homes.

Initial agreements between the State Department and the resettlement agencies involved sending 5,255 to California, 4,481 to Texas, 1,800 to Oklahoma, 1,679 to Washington, 1,610 to Arizona, and hundreds more to almost every state. North Dakota will get at least 49 refugees. Mississippi and Alabama will get at least 10.

Where the refugees go from there is up to the resettlement agencies in each state. Sometimes, refugees will ask to live in communities where they already have family or friends. But officials said that many of the displaced Afghans who arrived this summer had no connection to the United States.

“These folks are coming at a time when the job market is very good,” said Jack Markell, the former Democratic governor of Delaware who is overseeing the resettlement effort. “But they’re also coming here at a time when the housing market is very tight.”

“Our job is to provide a safe and dignified welcome and to set people up for long-term success,” he said. “And that means doing everything we can to get them to the places where it’s affordable, where we connect them with jobs.”

For Mr. Biden, failure to integrate the refugees successfully could play into the hands of conservatives who oppose immigration — even for those who helped the Americans during the war — and claim the Afghans will rob Americans of jobs and bring the threat of crime into communities. After initially welcoming the refugees, the Republican governor of North Dakota has taken a harder line, echoing concerns of his party about vetting them.

Haomyyn Karimi, a former refugee who has been a baker at an Afghan market in Little Kabul for thirty years, choked up at the thought of another generation of Afghan refugees struggling to build a new life in the face of financial difficulty and discrimination.

“They had lives in Afghanistan,” Mr. Karimi said through an interpreter during a brief interview at the Maiwand Market in downtown Fremont. “Their money was in banks in Afghanistan that are no longer available to them. So they’re literally starting with nothing.”

‘They need to find workers.’

The refugees are arriving at a moment of severe economic need — labor shortages across the country mean that communities are desperate for workers. In Fargo, where the unemployment rate is 2.8 percent, many restaurants have to close early because they can’t find enough workers.

“Everybody’s looking for people,” said Daniel Hannaher, the director of the Fargo resettlement office for the Lutheran Immigration and Refugee Service, which expects to receive several dozen refugees soon. “And, you know, it’s getting to the point now where everybody’s mad about the restaurants.”

The same is true in Tulsa, where the unemployment rate is 3.5 percent and dropping. G.T. Bynum, the city’s Republican mayor, told Public Radio Tulsa that he’s eager for the new refugees to see that Tulsa “is a city where we help each other out, whether you’ve lived here your whole life or you just got off the plane from Afghanistan.”

Financial help for the Afghan refugees flows through the resettlement agencies in the form of a one-time payment of up to $1,225 per person for food assistance, rent, furniture and a very small amount of spending money. An additional $1,050 per person is sent to resettlement agencies to provide English classes and other services.

Because refugees are authorized to work in the United States, much of the help is directed toward helping them find a job, Mr. Markell said. Refugees are also eligible to receive Medicaid benefits and food stamps.

Historically, refugees have quickly gotten to work in the U.S., without taking jobs from Americans.

About one in five new refugees to the United States finds employment in the first year of arrival in the country, a high rate among wealthy nations, according to a paper published by a trio of researchers at University College London last year in the Journal of Economic Perspectives. Employment rates for refugees to America jump sharply in the years that follow.

Critics of high levels of refugee acceptance, including top officials in the White House under former President Donald J. Trump, contend that refugees compete with American workers — particularly for low-wage jobs — and dramatically reduce how much those existing workers earn.

The vast majority of empirical economic research finds that isn’t true. An exhaustive report published by the office of the chief economist at the State Department examined settlement patterns of past refugees to the United States, comparing the economic outcomes of areas where they did and did not settle. It found “robust causal evidence that there is no adverse long-term impact of refugees on the U.S. labor market.”

If anything, economists say, the current labor market makes it even less likely that refugees would steal jobs or suppress wages for people already here. U.S. employers reported more than 10 million job openings nationwide in August, down slightly from a record 11 million in July. Workers have been slow to return to jobs or industries they left in the pandemic, leaving many restaurants and retail stores desperate to hire.

Few, if any, previous waves of refugees have entered the country with such high labor demand across the country, or with the lure of worker-parched areas that could offer relatively high starting salaries for even inexperienced staff.

And places like Fargo and Tulsa offer cheaper housing, too. The average rent for a one-bedroom apartment in Fargo is $730 a month, less than a third of what it is in Fremont. The average rent in Tulsa is $760.

‘Support is critical’

But some have concerns about sending the Afghans to places where there are few familiar faces and prejudice is more common.

In Michigan, which is slated to get at least 1,280 refugees, stickers with the racist message “Afghan Refugee Hunting Permits” were posted in Ann Arbor by the Proud Boys, a white supremacist group.

In Oklahoma, John Bennett, the chairman of the state Republican Party, posted a Facebook video in which he rants about the dangers of Shariah, the Islamic legal code, accusing the refugees — without evidence — of being terrorists.

“Oklahomans, I encourage you to call and email the governor, call and email your legislators, and tell them: Do not allow Afghan refugees into Oklahoma,” Mr. Bennett said in the video.

“We’re going to see Islamophobia. We’re going to see xenophobia,” said Spojmie Nasiri, an immigration lawyer of Afghan descent who lives near Fremont. “We’re already seeing it.”

But Mr. Markell said most communities — including conservative, Republican-leaning ones — have been very welcoming. He credits the country’s veterans, who have overwhelmingly embraced the Afghans.

“When they are as vocal as they have been, it helps a lot with elected officials of both parties,” Mr. Markell said.

Advocates say that despite having a higher cost of living and fewer available jobs, established Afghan enclaves like Fremont can provide a much-needed support network.

The International Rescue Committee, which operates a resettlement office in Oakland, Calif., near Fremont, said it had established committees on housing, health, case management and legal issues even before the mass evacuation from Kabul this summer. The Oakland office is expecting at least 600 to 700 Afghan refugees to be resettled in the area.

Those who go to Fremont will find a raft of existing services thanks to the presence of an estimated 25,000 to 30,000 Afghans in the city: adult schools to teach them English; mental health services aimed at people from Afghanistan; and informal help from area mosques.

Some local banks in Fremont are partnering with the city to provide financial coaching.

“That support is critical,” said Jordane Tofighi, the director of the Oakland office. “Some of the local mosques are doing food distribution. Some of the grocery stores have food pickup hours.”

Fremont also boasts social service agencies, including the Afghan Coalition, which have been catering to the community’s Afghan residents for several decades. Mizgon Darby, who works for the organization, has been pressing the resettlement agencies, local governments and the state to provide more financial resources for the latest wave of refugees.

“The question is, in these different areas that they are being settled into, who is the designated agency that is helping them in those cases?” Ms. Darby said during an interview in her Fremont office recently. “Who’s going to navigate for them or help them navigate?”

Mr. Karimi, the baker at the Fremont market, said he's hopeful that the latest wave of refugees will find the support they need to thrive in their new country. He said people like himself owe it to the new arrivals to support them with jobs, money and encouragement.

“If they want my blood,” he said, pledging his help for the new arrivals as tears streamed down his face, “I will give them my blood.”

Michael D. Shear is a veteran White House correspondent and two-time Pulitzer Prize winner who was a member of team that won the Public Service Medal for Covid coverage in 2020. He is the co-author of “Border Wars: Inside Trump's Assault on Immigration.” More about Michael D. Shear

Jim Tankersley is a White House correspondent with a focus on economic policy. He has written for more than a decade in Washington about the decline of opportunity for American workers, and is the author of "The Riches of This Land: The Untold, True Story of America's Middle Class." More about Jim Tankersley

Advertisement

Yggdrasil Urban Wildlife Rescue (YUWR) and Education Center

Wild Emergency

• Skunk Problems
• Raccoon Problem?
• Fawn Stories
• Raccoon Rehabs
• Dozer The Gulf Coast Box Turtle
• Two Orphaned Squirrels….
• Forever Homes Needed
• Successful Adoption Stories!

If you are having a problem with a wild animal invading your home or business and are looking for a humane alternative to a lethal Pest Company, please visit our * Wildlife Problem *  page for easy solutions.

If you have a wildlife emergency , as in, an injured or orphaned wild animal, please read the emergency information below. if you still need help, read our contact information below..

To report poaching ( intentional killing or injuring of wildlife ), polluting of habitat, etc. please call the CA. Dept. of Fish & Wildlife Hotline at (888) DFG-CALTIP

• I HAVE A WILDLIFE EMERGENCY – WHAT DO I DO?
• WHO TO CALL

Raising an Orphaned Animal

I found a baby songbird. What do I do?

• I found a baby HUMMINGBIRD
• I found a baby Jackrabbit
• I found a baby Deer (Fawn)…
• I found an Injured Adult Deer…
• I found a baby Squirrel…
• I found a baby Opossum…
• I found a baby Raccoon…

Wildlife Emergency – What do I do?

If you have found an injured or orphaned wild animal:

• Please read this page to determine if the animal is actually in need of help.
• Do not give the animal any food or water. A captured animal will get food and water stuck in its fur/feathers potentially leading to discomfort and hypothermia. Feeding an animal an incorrect diet can result in gastrointestinal injury or death.
• Keep the animal warm. Place in a box with a towel, tissue, or old t-shirt. Make sure the box has holes in the lid. If you have a heatingpad, place the box halfway on the heatingpad and set the heatingpad to LOW only.
• Keep the animal in a dark, quiet place.
• Leave the animal alone; don’t handle or bother it. Keep children and pets away to avoid death by Capture Myopathy.

If the animal you have found is a bat, skunk, fox or coyote, do not handle the animal with bare hands. Because these animals are rabies-vector species, they MUST be handled with gloves or towels !  It is a strict rule in California that if someone touches a rabies-vector animal directly, the animal MUST be euthanized and tested for rabies to ensure the health and safety of the finder and the community.

back to top

Who To Call:

Yggdrasil urban wildlife rescue – (510) 421-9897:, we only take infant mammals, at the present time, due to loss of our main facility, we can only accept  baby mammals. please do not drop animals off at our previous locations as we are no longer there and the animals left there will die . to arrange drop-off please call or text us first at 510-421-9897. thank you., for birds and adult & injured wildlife mammals  in the san francisco bay area:.

• Lindsay Wildlife Museum in Walnut Creek  (NATIVES ONLY – NO PIGEONS or SKUNKS) – (925) 935-1978
• Sulpher Creek in Hayward – (510) 881-6747
• Ohlone Wildlife in Fremont – (510) 797-9449 (PIGEONS OK)
• WildCare in San Rafael – (415) 456-SAVE (PIGEONS OK)
• Peninsula Humane in Burlingame – (650) 340-7022
• San Francisco Rescue Orphaned Mammal Program (SFROMP) in San Francisco – (415) 350-WILD or (415) 554-9400
• Or look up your nearest Wildlife Center

FOR WATER BIRDS :

Contact the international bird rescue and research center wildlife hospital at: (707) 207-0380 ext. 110, wildlife emergencies outside the sf bay area :, for  all other areas , please  use the following links to find the wildlife center nearest you:.

• In California:  To find a wildlife rehabilitation center nearest you in California, visit   this    List of California Wildlife Rehabilitation Centers
• Nationally: Here is a great website with resources on finding a wildlife center anywhere in the United States:   http://www.nwrawildlife.org/?page=Find_A_Rehabilitator .
• Internationally : To search Internationally, go to  https://theiwrc.org/resources/emergency

For many bird species, leaving the nest before being flighted is a natural part of adolescence. Young birds with developing feathers frequently take up residence on the ground in the grass or bushes near their old nest where they continue to be fed and taught by their parents. Some things to keep in mind about birds:

* Parent birds will continue to feed their babies after you have touched them. Most birds have a poor sense of smell and parent birds won’t know you have touched their baby. Baby birds can be returned to the nest! Parent birds will even be foster parents for an abandoned baby of the same species and age of their own young.

* Parent birds will search for their babies even after 24 to 48 hours of absence. Most birds have their own territories. Even if the nest and babies are gone, the parents remain in their home territory, waiting to welcome their babies home.

* Birds only need to be rescued if they are:

• caught by a cat or dog
• naked (no feathers)

* Symptoms of an Injury or Illness:

• Falling over on one side
• Unable to flutter wings
• Weak or shivering
• Attacked by cat or dog
• Wing tweaked upward
• Wing drooping
• Feathers fluffed

Has the Bird Been Abandoned? Watch for the parents… observe the baby bird continuously for 60 to 90 minutes from a distance of 50 feet. Watch carefully; the parents will fly in and out quickly.

If you have the bird in a box, check the feces…

• Clear with white poop (or green bile) indicates a baby bird is not being fed, and is likely abandoned.
• Color in the poop indicates that the parents are feeding the baby, and the bird should be put back where it was found.

Fallen Babies: Naked and pin-feathered birds should be kept warm while trying to locate their nest. The babies will get chilled quickly.

An entire nest of birds can be placed in a small tissue-filled wicker basket or butter tub with drainage holes in the bottom. Nail the basket or tub to a tree in a location safe from crows and hawks (with some tree-cover). Be sure that a branch shields the nestlings from sunburn.

One single baby must be returned to the original nest with its siblings. Parent birds will only sit on and feed the babies in one nest. If the nestling cannot be returned, contact your local wildlife rehabilitation center.

Fledglings:

These birds have feathers and short tails and can perch, hop or walk. They are learning to fly, a process that may take two weeks. Fledglings should be left alone to practice hopping and fluttering from low shrub branches to the ground. The parents are close by, and continue to feed the babies until they learn to fly and eat on their own. Parents will guide the fledglings into the bushes at night to hide from predators.

To return a fledgling to its territory after it has been brought indoors:

• Keep pets and children indoors so the parents will return to their baby.
• If a bird can perch on your finger, place it in a bush near the area you found it.
• If you found the bird in a high-traffic area, move it to a nearby safe area under the cover of bushes. Don’t worry – Parents communicate with their young by a series of voice calls. As long as a fledgling is placed in its home territory, the parents will be able to locate it and move it to a safe location.
• At a distance, (indoors is best) watch continuously for one hour for the parents to return. If the parents don’t return, call your local wildlife rehabilitation center immediately.

I found a Hummingbird. How can I help it?

If you find an injured or orphaned hummingbird on the ground, lift it along with the material it is sitting on, and place it on crumpled tissue in a shoebox with holes in the lid. Always use tissue or paper towels, NOT cloth—the bird’s feet may become entangled in the cloth.

Call your local wildlife rehabilitation center immediately. Hummingbirds will die within four hours if not fed. Hummingbird babies that are fed sugarwater or commercial hummingbird nectar for more than 24 hours may develop crippling deformities.

Never attempt to remove baby hummingbirds from their nest. Young hummingbirds secure themselves to the nest by weaving their tiny toes around the nest fabric. So firm is their hold, that if lifted from the nest, most often the legs are left behind.

I found a baby jackrabbit. Is it orphaned?

It can be very difficult to tell if a baby jackrabbit is orphaned. Some facts to keep in mind about jackrabbit behavior:

* Jackrabbits are “precocial,” meaning they are born fully furred with their eyes open, and will start nibbling greens within a week. * A baby jackrabbit is called a leveret. A leveret’s main defense when threatened is to freeze, which is often mistaken by people as being calm. The animal is not calm, it is terrified. * The mother jackrabbit separates her litter for a better chance of some babies surviving. * There is no nest! Jackrabbit young will stay hidden in the grass, shrubs, or other ground-level growth where the mother leaves them. * The mother only comes to feed the young two or three times a day. Otherwise, the young jackrabbits are left alone. Even if you are watching carefully, you may not see the mother jackrabbit return to her young. Do not assume the babies are orphaned simply because you do not see the mother! * Leverets will wander a bit. When the mother returns to the area, she calls to her young and they come to nurse. * Never try to feed a leveret, they have very delicate digestive systems. * Jackrabbits are extremely high-stress animals; they can die from fear. * Jackrabbits have very strong hind limbs and if restrained may kick out hard enough to break their own backs.

I found a baby deer (fawn). What do I do?

* The mother only comes to feed the young two or three times a day. Otherwise, the young fawns are left alone while their mother is out searching for food to make more milk. Even if you are watching carefully, you may not see the mother return to her young. Do not assume the baby(s) are orphaned simply because you do not see the mother! Fawns will wander a bit. When the mother returns to the area, she calls to her young and they come to nurse.

If a fawn is seen lying upright, eyes wide open, but flattened to the ground, do not touch it. This is a fawn’s camouflage position. It blends with its surroundings. When it is picked up it will hold its legs tight against its body with its head forward. Its legs are not broken. Sometimes the fawn allows its body to become limp and dangle in your hands. Put it down, walk away and leave it alone. This fawn is too small to follow the doe for the long distance she must travel to find enough food to make milk for her baby. A doe may leave her baby alone for up to 6 hours at a time in her search to find food. Doe’s milk is very rich and will sustain the fawn for the many hours it spends alone. The doe will return only when there are no humans nearby. Do not sit and wait for her to return. If you have removed the fawn from its resting spot take it back at once and walk away. The doe will be searching for her fawn.

* Deer are extremely high-stress animals; they can die from fear. Petting the fawn, talking to it, holding it, does not comfort it. This is a wild animal. Human voices, odor and touch only add to the stress and will cause additional harm besides the illness or injury. When a fawn seems calm it may be in shock. (Read about “Capture Myopathy” here)

If a fawn is obviously ill, lying on its side, kicking, crying – pick it up and place it in a quiet place. A light cloth placed over the animal’s head will sometimes calm it. Keep it away from pets and all human activity.  If the weather is cold, a blanket may be placed over its body to keep it from becoming chilled. In hot weather a cool location, out of drafts, and call your local wildlife center.

*Deer have sensitive digestive tracts. DO NOT FEED THE FAWN ANYTHING other than water. Baby formula, cow’s milk, feed store mixes, pet store domestic animal formulas, soy products – will cause scouring, dehydration and death. CALL A WILDLIFE CENTER at once for help.

I found an injured ADULT Deer. What do I do?

Unfortunately there are no resources to help injured adult Deer, anywhere. The good news is that if the animal is mobile, the ability to heal and survive is tremendous. I have seen deer with a leg dangling off heal up and survive for years!   Please read THIS article  for more tips on how you can help injured adult deer.

If the animal is NOT MOBILE, is laying prone, has a broken back, hip, or head injury, and is suffering with no chance of recovering on it’s own, then helping by contacting agencies that can access and provide a swift end is the most merciful option, rather than letting it linger and suffer for days before it finally dies on it’s own. There are some services for this and they mostly involve calling your local animal control agency who will come out to assess and if they observe that the deer is not savable, they will call the local police department to dispatch it quickly. Here are some numbers to help you if this is the issue:

• In Berkeley, please call BACS at 510-981-6600
• In Oakland, please call 510-535-5602 or after hours, 510.777.3333

While the deer is laying helpless in your backyard, I am sure you will have the impulse to try to comfort it. This is normal. However, because Deer are wild animals, they react to your petting and comforting differently than you might intend. Please read below about Capture Myopathy to understand better how our attempts to comfort may make the deer suffer more than if we were to just back off and leave it alone while we wait for help to come.

“Capture Myopathy” – What is it?

Capture myopathy (or white muscle disease) is a response by prey animals to abnormal stressors. It is a syndrome of acute muscle degradation resulting from stress, especially fear. It can occur without exercise (animal does not have to be chased or even captured).

Capture Myopathy occurs in Prey animals such as deer and rabbits and it is Mother Nature’s built-in defense mechanism for situations where these animals have been captured by predators. Once captured, to prevent them from suffering while being eaten alive by a hawk or cougar, their body releases hormones that break down the muscle tissue in the heart, resulting in death.

The clinical signs of capture myopathy include sudden death within 24 hours, depression, rapid shallow breathing, and failure to recover from anesthesia. Death can occur after several hours of symptoms, or from cardiac arrest. The animal may also appear to recover, but has heart damage. It may die at the next stressful event.

There is no treatment for capture myopathy. Prevention is the only treatment. That’s why you’ll get the same advice from rehabbers across the country. Keep the animal dark and quiet until you can get it to help. This reduces stress and the therefore the chance of capture myopathy.

I found a baby squirrel. What do I do?

Loud Chirping? Click here to hear the sound of a baby squirrel in distress

Making it possible for Mom to get her baby back is the very best thing that you can do. Baby squirrels fall out of trees all of the time – whether they are learning to climb, playing with siblings, or just hit with an unexpected brisk wind. If alive and healthy, Mom will most definitely try to retrieve her baby if her baby is warm and healthy, regardless of whether or not you have touched it. Here is a CHECKLIST to see if the baby is able to be reunited with the mother:

• Is the baby visibly injured or badly bleeding?  (Bloody noses are common and are not necessarily a sign of bad injury- mom will often still retrieve them.)
• Is it very cold to the touch? (Baby squirrels should feel warm.)
• Is there a dead adult nearby, that you are SURE is the parent? (If you are only guessing, try the reunite still.)
• Is the baby very skinny or wrinkly? This could mean that the mother has been missing for a long time and the baby is starving and dehydrated.
• Are there flies or maggots on the baby or in any of its orifices?

If you answered YES to ANY of these questions, the baby is NOT reunitable and you should contact us immediately via call or text at 510-421-9897.

If you answered NO to ALL of these questions, please use the following instructions to attempt to reunite the baby(s):

HOW TO REUNITE

Please start trying as soon as you can or as early in the morning once the sun is up as this is an active time for squirrels. Always bring babies inside at nightfall , if mama has not retrieved them yet, and contact us.

• If you have found a single baby check the area for any siblings.
• Place the baby(s) in a small box surrounded by soft material (fleece, an old t-shirt, or the original nest), with instant hand warmers or a tightly sealed hot water bottle.
• Place the box near where the baby was found, preferably mounted as high as possible off the ground (bungee cord or rope can be used to affix the box to a tree).
• Place a bluetooth speaker inside the box with the babies – making sure the speaker sound is facing away from the babies so it doesn’t hurt their ears. If you do not have a bluetooth speaker you can use your cell phone.
• Using your connected bluetooth device, please visit this youtube page:  https://www.youtube.com/watch?v=tgXnLN9w5BA   and play this over the speakers so the mother can pinpoint on where her babies are. She should come and retrieve her babies before the 1 hour video is over. If not, please repeat.
• Leave the immediate area and observe from a distance- a mother squirrel won’t return if people are nearby or chainsaws are still being used nearby. The mother squirrel should return to retrieve her baby within a few hours. Mother squirrels usually keep several nests, so it’s not a problem if it was destroyed, they will simply take their babies to a different location.

If the mother squirrel still does not return by dark, retrieve the baby, keep it in a warm place, and contact us at 510-421-9897. Do not try to feed or give water to the baby squirrel. Baby squirrels require a very specific diet and feeding a baby squirrel the wrong food or without the proper supplies and training can make them very sick and be fatal to the baby. (bubbles coming out the nose means the fluid went into the lungs because they were fed incorrectly and they can die from this, for example) Please don’t feed them. Call us first. If you want to join our fostercare team we would love to have you and will teach you how to do everything right so the babies grow up strong and healthy.

I found an opossum. What do I do?

Opossums are marsupials. Their babies are born very tiny, no bigger than a dime, and the mothers carry them in a pouch. When the babies grow bigger, they cling to the mother’s back until they are so large they fall off and start life on their own. Solitary, nomadic, and nocturnal, they are rarely active during daylight; they look for a safe, dark place to sleep. Opossums eat a wide variety of foods, including garden snails, slugs, insects, mice, fruits, and snakes. They will also opportunistically seek out dog and cat food, as well as trash. When confronted or startled, they will run away or may “play dead”—lying motionless for up to an hour. If cornered, opossums gape their mouths, showing off their 50 pointy teeth, trying to look as fearsome as possible.

Common problems:

Hit by car or attacked by dog : If it can be safely put into a secure container without touching the animal, it can be brought to the wildlife hospital or call Animal Services. If you find a dead opossum, check to see if it has a pouch on the lower abdomen. If there are babies in the pouch, bring them to a wildlife hospital

Orphaned young opossums : If the opossum is 8 inches from the nose to the base of the tail or larger, it is okay to leave it alone—this is the normal age for them to leave their mothers. If it is less than 8 inches or is injured, place it in a warm, dark container and bring it to a wildlife hospital as soon as possible. They require special diets, so no food or water should be offered.

I found a raccoon. What do I do?

Raccoons are nocturnal—they need a quiet, dark place during the daytime. They are attracted by pet food, ripe fruit, and water. Their nesting season begins in February and can go through October. Litters average 2 to 7 babies, and they are weaned after 12 to 16 weeks. Adult raccoons are usually solitary, but young raccoons may stay with their mother during the winter, either in the same den or nearby. They den in tree cavities, underground burrows made by other animals, and human-made structures, such as chimneys, basements, attics, spaces, under patios, and between walls.

It is not good to feed raccoons, either intentionally or unintentionally. Don’t leave pet food outside at night; keep ripe fruit picked; and keep garbage cans securely covered or inside a garage at night. Raccoons are excellent climbers and are very dexterous. Raccoons cannot be relocated. A territory left open by removing one animal will be quickly filled by another. It is also illegal in California to relocate animals. Please do not give food or water to any injured or orphaned animals.

Young raccoon alone : If under the house or in a nest area, leave them alone so the mother will return. If in an inappropriate area (out in the open), wait until evening to see if the mother returns. They may be placed in a box that they can’t get out of, but the mother can get in. Do not handle the raccoons with your bare hands. Since raccoons are nocturnal, the mother will not return until night. If the mother doesn’t return, bring them to a wildlife hospital.

Hit by car : If the raccoon can safely be put into a secure container without touching the animal, it can be brought to the hospital. Otherwise, call animal control, especially for adult raccoons.

Please do NOT leave a comment here with a wildlife emergency. No one will read it in time. Please use the information above to solve your problem or call the appropriate place for help. Thank you.

A Nonprofit You Can Trust

VISIT US ON FACEBOOK AT: www.facebook.com/YUWR.org/    for our most recent wildlife tales

Recent Posts

• 2023 In A Nutshell – Our End-Of-Year Review December 29, 2023
• Our Live Facebook Feed 2023! January 10, 2023

Live Air Quality at our San Francisco Intake Center

Live temperature at our san francisco intake center.