The use of UAV swarms is becoming more widespread due to the reduced costs of UAVs and their ability to accomplish tasks more quickly and effectively as a group rather than individually. Advancements in aircraft design and control, communication topologies, and battery systems have made coordinated UAV swarms possible. The use of UAV swarms in applications of military [1, 2], ecology [3], remote sensing [4, 5], disaster management [6], crowd control, emergency communication [7], agriculture [8, 9], and victim search [10, 11] are just some of the use cases. As individual and multi-robotic development and their interaction with real-world entities advances, these applications are only limited because of a lack of improvement, a discrepancy that exponentially decreases as time passes. With the increased diversity of swarm usage applications, research in UAV resiliency has also grown [12, 13]. Due to the close-knit topology of these swarms, the failure of agents above a certain threshold can often lead to cascaded systemic collapse and a pause on mission progress. The cause of this failure may be structural, such as in a leader-follower topology, wherein the followers may get disconnected if the leader fails. Additional uncertainty in this failure also exists, such as the possibility of a failing swarm agent crashing into other agents during its collapse. Resiliency is defined as the ability of a system to withstand disruptions. Broader definitions include the ability of a system to bounce back after a disruption [14]. D.D. Woods summarizes system resilience perfectly in its four core concepts [15]. They are resilience, such as rebound, robustness, extensibility, and adaptability. Previous work by the authors addresses systemic resilience in UAV swarms on a broader range by classifying UAV swarm operations into components and modules [16]. The resilience of UAV swarms is a complex topic that integrates multiple components of navigation [17], mapping [4], control [18], defensive and intrusion detection policies [19], agent welfare, and physical characteristics of the swarm agent [20] into an intricate system designed to create balance in a dynamic environment.

An MRS (Multi-Robot System) and a swarm are both concepts in robotics that involve the coordination and interaction of multiple robots. However, they have distinct differences in organization, control, and behavior. An MRS has a structured and explicit interaction scheme with a centralized control. Swarm agents, on the other hand, are more decentralized and self-organized. There might not be explicit communication between robots, and collective behavior emerges from simple interactions between individual robots following local rules.

Additionally, MRS may involve centralized controllers and planners to assign tasks to robots. Swarms rely on local interactions and distributed control. Each robot can typically make decisions based on its immediate surroundings or information gathered from nearby agents. This also involves a degree of autonomy in decision-making from completely autonomous to semi-autonomous. However, while swarms are expected to be inherently scalable, there is an ongoing debate on the minimum number of agents that must be present and acting collectively to label it as a swarm. Adding more robots to a swarm does not necessarily increase performance. Selecting the number of agents in a swarm has long been contested. However, approaches with agents as few as five and as many as 1,000 have been implemented and studied. Article [21] discusses how aspects such as system scalability, technical capabilities of individual agents, and financial or logical constraints influence the selection of the number of agents in the swarm.

These factors were crucial in selecting the number of agents for swarm response experiments performed in this study. The number of agents available for experimentation was limited. Additionally, some agents were designated as reserve and spare agents to ensure experiments continued in case of equipment failure. Space constraints allowed only a certain number of agents to fly in the designated airspace without the risk of agent collision and crashes due to induced airflow interactions. While all communication between agents was performed decentralized, primary communication and network protocols required for communication with GCS limited the number of agents connected to them.

Search and Rescue is a vast domain; focusing only on swarm agent welfare significantly narrows it. However, to concentrate results further, this SAR framework will be primarily described for exploratory swarm applications. Scenarios where a swarm of agents may be deployed over an area and, in the process, may lose contact with the swarm is the priority. This narrows down the framework focus as well as experiment design and validation. Two major types of SAR capabilities in UAV swarms are defined and categorized here. They are application-specific SAR [22] and swarm-specific SAR (SS-SAR). Although our study takes a different direction than a regular application-focused SAR use scenario, it remains an exploratory problem. The tracking, location, and Rescue of disabled swarm agents require other agents of the swarm to actively search the target space for the agent using techniques such as triangulated localization, computer vision, sensors on the ground, and the analysis of system-generated mission logs. A literature review reveals that most swarms lack the self-awareness needed to actively take care of their agents. More robust mechanisms for the welfare of UAV swarms [23] are needed as an additional means to increase systemic resilience.

To build robust applications and routine case scenarios that use UAV swarms, the swarm itself must be resilient to disruptions. Towards this goal concerning SAR swarm operations, the significant contributions of this paper are:

- A literature analysis that reveals a research gap in UAV swarm development related to the search and rescue of their agents.

Using a decentralized approach for communication and agent decisions, this study aims to demonstrate scalable and robust responses of swarms to disruptive scenarios along with further scope for possible emergent behavior to avoid them altogether based on broad programmed constraints. The paper is arranged in the following way. Section Introduction gives a brief introduction of the area with research contributions of this study. Section Summary of Recent Literature on Multi-Robot SAR presents a categorization of current trends. Section Swarm-Specific SAR Framework presents the SS-SAR scenario description and framework workflow. Section Performance Tests describes both hardware and simulation performance tests and environmental parameters. Section Results presents the experimental results of the new SS-SAR framework. Section Discussion and Future Research Directions provides future directions to approach the problem with suggested framework upgrades, and Section Conclusion provides concluding statements.

The multi-UAV SAR problem is a broad problem domain. This section categorizes current research into three distinct approaches. Table 1 identifies research on the topic and categorizes the study as application or swarm-specific problems. Application-specific SAR (AS-SAR) and SS-SAR are the two categorizations previously discussed. Depending on how the SAR problem is approached, a third category is also included: Search Methodology Focused (SMF). Research in this category does not have a specific search target type; instead, the focus is on the general SAR methodology, where any internal or external search target can be assigned.

As one can see from Table 1, swarm-specific implementations are less explored in the literature. In addition to the above literature review, generalized methodologies exist that propose novel approaches that would improve facets of the SAR process. These include using bio-inspired algorithms for area coverage [42] formation tracking [43] and environment exploration [44], updated and merged observation maps or information exchange pathways [45], and efficient task planning [46, 47]. Frameworks such as [41] that propose automatic replacement of lost UAV agents are scarce. This example fits perfectly in this paper’s proposed swarm-specific research category. To keep the literature analysis attainable, any approaches that do not directly describe the use of aerial vehicle swarms in the field for SAR have been eliminated. This includes broader research topics such as using machine learning methods to improve object recognition in aerial images taken by UAVs [48].

Hardware and software tests were designed to test the proposed workflow under different conditions. Experiment range and series were selected considering the range and variability required to effectively demonstrate performance [53]. Table 3 summarizes the primary objectives of each test, the map used, and the number of experimental flights performed. Overall, these tests represent a modular approach to developing and testing an SS-SAR framework for increasing the operational resiliency of a UAV swarm system.

Each performance test was associated with a map, as summarized in Table 3. The hardware tests were performed in maps M1 and M2, and the simulation tests were performed in maps M3 and M4. Table 4 outlines characteristics of the map environments used in the performance tests.

The proposed framework is quite flexible regarding the agents that can be used. However, at minimum, lateral and downward vision sensors are required, along with either a GNSS module or capability for passive beacon georeferencing. Considering hardware and fly space limitations, an indoor location was used for hardware tests. The DJI EDU [50] UAV platform was chosen to perform hardware performance tests. These low-cost drones provide a basic environment for drone testing and flights. In the past, there have been multiple approaches to using Tello drones as platforms for singular and swarm development. The authors of [54] use Tello agents to demonstrate an automated swarm flight in a restricted flight space. A matrix formation control that uses Tello to display patterns was adopted in [55]. In [56], the DJI frame was used to build visionless sensing drones for obstacle avoidance and maze solving.

Related research such as this assisted in realizing the various limitations of the Tello platform during experiment design. The Tello agents are low-cost entry-level hardware and are intended for proof-of-concept experiments. Without a dedicated GNSS receiver, the agents rely on a VPS using the downward-facing camera module to localize using ground planes and additional GCPs. All recorded video and image data is streamed in real-time to ground control without storage and post-processing ability. While these constraints prevent executing a full-scale framework representation on these agents, our experiments modify the complete framework based on its modular structure. This modular and stepwise process permits testing smaller decision statements using simple Tello agents. Table 5 contains manufacturer-provided specifications for the DJI drones. These specifications have been referenced from online user manuals [50].

The following section highlights observations recorded during each performance test and their analysis. Figure 15 shows PT1 time to distress logs, the time when the agent first experienced an issue, and the time to rescue, which is the amount of time the rescue agent took to move to the position of the fallen agent and rescue it. Out of 15 flights, the rescue agent successfully rescues the distressed Tello agent nine times, as denoted by a green dot in Figure 15. A preliminary p ₂ check was performed using the available battery percentage when rcm was successful. The battery values are in Figure 16. If the battery value after successful reconnection was established between the Rescue and the distressed agent was below the given threshold (50%), the distressed agent was deemed incapable of rejoining the swarm. This was observed during flights 5, 7, 12, and 14. Each flight was independent, and the battery was charged to maximum capacity before each flight. Collision occurrence was counted when the rescue agent experienced collision at any time during the rescue process. As such, those flights were recorded as an unsuccessful recovery. During flights 4 and 10, the rescue agent experienced a collision and could not recover the fallen agent successfully; these flights were logged as failures. It was observed that close interaction with swarm agents in the constrained airspace caused unpredictable drifts in agent movement due to induced airflow, resulting in collisions and crashes. Overall, a recovery rate of 60% was thus calculated for PT1.

Figure 17 shows PT2 performed on M2. For ten flights, p1 and p ₂ rescue decisions were recorded. This test aimed to observe these pose checks and how they affect agent recovery. Flights 4, 7, 8, and 9 showed a successful recovery. Flights 1 to 3 and 5 failed p _1, where it was determined by the operator using the rescue agent that the fallen agent was not in a position from which it could safely take off. In M2, this scenario was due to indoor obstacles, such as furniture, that might prevent the agents’ safe take-off ability. In Flight 6, the agent passed p _1, which denoted it was in an orientation and position that could enable safe take-off; however, it failed real-time pose check p ₂ . For flight 10, both p₁ and p ₂ were successful. However, the agent could not take off due to an internal malfunction. A success rate of 40% for PT2 was observed.

In addition to previous pose checks, PT3 on M3 also performed preliminary pose check p ₀ on the agents using the distressed agent’s onboard vision sensor, as shown in Figure 12. Figure 18 presents the ten flights performed. Flights 3, 5, 6, and 8 successfully recovered the distressed agent. For flights 1, 7, and 10, the agent failed p ₀ , indicating that the agent was not in a position to take off safely. As a result of the preliminary pose check failing, the rescue agent was not deployed to conduct further observations. In Flight 2, p ₀ passed. However, p ₁ failed, leading to a failed rescue attempt. In Flights 4 and 9, a successful p ₀ and p ₁ were observed. However, the agent failed real-time check p ₂ and was thus labeled unrecoverable. A success rate of 40% was observed for this test.

PT4 on M4 further examined an additional ten flights, and the results are shown in Figure 19, where time to distress logs the time a swarm agent experiences an issue, and time to rescue logs the time the rescue agent takes during rescue attempts. Flights 4, 5, 6, 7, and 9 showed the rescue agent’s successful recovery of the distressed agent.

Further examination of operational parameters, as shown in Figure 20, gives additional failure information. In flights 1, 8, and 10, the distressed agent passed p _0, which denoted its orientation passed requirements for safe rejoin. However, p ₁ failed. Since PT4 was performed on an outdoor terrain map that included tree obstacles, the primary reason for p ₁ to fail was the tress obstructing safe rejoin maneuvers. In-flight 2, the distressed agent failed to p0 itself, as denoted by the onboard sensor that gave information regarding its orientation and crash position. The floating window views in Figure 14 for the distressed agent one vision sensor FOV show an example of an agent that has landed upside down. The inverted image shows the ground above and tree foliage below. In flight 3, the distressed agent passed all required pose checks; however, a collision with the rescue agent resulted in a failed attempt. Overall, a success rate of 50% was observed. A summary of success rates for all performance tests is shown in Figure 21.

A test on a sparse block map depicted in Figure 22 was used to log agent loss and recovery incidents. Every distress signal was mapped as a triangle or square in that process. The triangles indicate agent loss due to network issues, and the squares indicate agent loss due to collision. Red shapes indicate an unsuccessful attempt at recovery, and blue shapes indicate successful agent recovery. Each shape is the result of a separate flight, and consecutive flights did not have prior risk zone information. However, creating such risk zones can then be used as future input parameters to create safe flight paths. For example, if an area sees increased collisions due to dense obstacle geometry, a threat area can be modeled where agents entering that area do not venture below a preset altitude to avoid collisions. If agents moving to a particular area lose connection with ground control, the next iteration framework run will adjust the upper bound distance between the agents, which defines the maximum distance between two agents based on SSI _n . Adjusting the upper bounds will result in agents flying in close formation and using data hop pathways to connect to ground control and prevent agent loss due to network range limitations. Future work using this approach can demonstrate adaptability, robustness, and emergent behavior in the swarm based on simple governance rules.

The proposed framework is a preliminary step in developing robust methodologies for evaluating swarm awareness toward the wellbeing of its constituent agents. This includes testing capabilities such as keeping track of each agent’s progress toward its task, realizing the occurrence of agents in distress, locating the distressed agents, and initiating rescue operations to enable them to rejoin the swarm. Several modifications could be implemented via the modular nature of the designed framework, as initiated by research directions summarized below.

It is crucial to consider the impact of emerging regulations on UAV operations, particularly the recent implementation of the FAA’s Remote Identification (RID) rule [58]. This regulation mandates the use of Remote Identification modules on certain UAVs, allowing for the open broadcast and identification of these agents during flight. This rule ensures safer airspace and promotes regulated use of UAVs, UAV swarms, and their applications [59]. When integrated into our rescue framework, the potential for such information can significantly enhance tracking and rescue performance. By leveraging the real-time identification capabilities provided by RID, it is foreseen that such frameworks can precisely locate and rescue other agents within the swarm more effectively, thus bolstering the overall efficiency and reliability of the proposed UAV rescue mechanism. Furthermore, exploring the compatibility and interoperability of our rescue system with other upcoming regulatory frameworks will be essential in ensuring the seamless integration and widespread adoption of our research in real-world UAV swarm applications.

Currently, some sections of the framework involve human decision-making. Most notable is the analysis of the distressed agent pose data transmitted by the rescue agent. The human operator observes the images to create a preliminary decision on the fallen agent’s possibility of rejoining the swarm. The human in-loop component can be reduced by adding autonomous UAV detection capability that uses vision sensor data, deep learning, and image processing techniques. This is possible using approaches such as [60] that use agent vision sensors for target analysis. An additional upgrade involves multiple agents to capture disruption and distressed agent information from different angles to gather a richer dataset.

A modified task re-allocation algorithm would enable additional agents to join the swarm and take up the interrupted task of the fallen agent or a re-allocation scheme for existing swarm agents to assume responsibility for the incomplete task. Resource allocation implementations such as in [61–63] exist that could be implemented. Further experiments could be explored for loss rates in the same airspace with LiDAR (for obstacle detection) and preset waypoints in a map (using GNSS). This would expand the feature of the existing framework to create probability maps. While current risk zones were labeled using agent failure location data obtained from transmitted HBS, future experiments may include the presence of VRRZ. This is the system’s ability to create variable radius risk zones. Each risk zone can have a variable radius, thus allowing the mapping of larger disruptive structures to be represented more accurately.

The above study uses agents with similar capabilities acting in the same operational space. Including a diverse range of agents identified by their differences in nature, hardware, or operational space introduces heterogeneity in the swarm. The impacts of such inclusion on the performance of SAR agents can also be explored. Some existing research investigates the possibility of using a swarm composed of heterogeneous agents for victim detection after a disaster [64]. Although their main goal was exploring how swarm heterogeneity can affect performance, they modeled a target search and rescue problem to study it. Their proposed technique differentiated between different agents and labeled them as heterogeneous using behavior trees. A positive correlation was produced between the swarm’s heterogenous capability and the time to search and rescue the target. A similar approach can be explored in the future, where differently abled robots are introduced in the swarm and are tasked with looking for swarm agents whose operations have been disrupted during mission progress. In the above experiments, all distressed agents were located on the ground. Thus, adding a UGV to track and locate the fallen agents to create an in-depth pose check analysis would be a logical step for further exploration. Several implementations of heterogeneous swarms exist, such as UAV-UGV collaboration [65], UAV-UWSV [29], and UAV-UGV-UWSV [66], demonstrating promise for more effective results than a single operational space swarm.

Intrusion detection systems can be implemented on the UAV network as a backend process. While IDSs are most prevalent on traditional networks to deter unwanted network access and activity, current lightweight versions have been shown to run reasonably well on MANET and FANET deployments with acceptable performance [19, 67–70]. Various types of IDS are available depending on their makeup and method of detecting malicious entities [71]. IDS could detect external agents attempting to maliciously disturb swarm operations. Similar approaches could also address ground-based attempts to take over swarm networks. The possible advantage would be the existing periodically transmitted HBS signal that can be used as input to any IDS. Adding network transmission data from each agent in the HBS could be used to design either a rule-based or anomaly-based lightweight IDS, at the very least. In this way, the SAR framework could provide additional security features to the swarm using inherently built structures.

A different approach taken to designing robust behavior was observed as a way of defining reliability in systems. The methods used in this category implement preemptive strategies for maintenance, abort policies, or recovery actions. This alternative form of resilience integration calls for an independent study. However, the results of the brief survey conducted on it are summarized below. These methods can be viewed as possible implementations and upgrades to this proposed SS-SAR framework. Table 8 summarizes the examined work based on their development focus being broader systemic implementations or UAV swarm-focused.

This research addresses gaps in current swarm resiliency research by addressing swarm-specific SAR rather than application-specific SAR. The approach was not to replace current SAR methodologies but to create an add-on that enables them to keep track of swarm agents while performing other functions. Modular experiments conducted on real-world hardware and simulations validated the need for, the possibility of, and the success rate of swarm-specific SAR approaches. While low-cost Tello drones were limited in their ability to handle a complete SS-SAR framework, they were crucial in testing the constituent process of the framework, such as reconnection protocols and pose check handlers. Simulation results provided a greater insight into how such frameworks can handle swarm agent loss. Experimental results prove that focusing on this approach to resiliency integration in multi-agent systems can produce the anticipated benefits. Recovery rates of distressed agents during and after the mission process increased drastically, especially in systems with no contingency rules. UAV swarms are complex and highly dynamic, making integrating resilience factors much more arduous. A system must exhibit awareness and diagnosis capability regarding its health before and after a disruption to efficiently produce solutions to mitigate said disruptions. This swarm-specific SAR framework is a crucial design step in that direction.