1 Introduction

In recent decades, single-UAV systems have been used for military purposes such as border search and surveillance, target detection and attack and also civilian applications including wildfire monitoring and firefighting, traffic monitoring and management, ad hoc networks and disaster monitoring [1]. However, in wide area applications, using a cooperative team of UAVs seems to be more efficient than a single UAV. Multi-UAV system has advantages over a single UAV which include scalability, lower cost, smaller radar cross-section, faster completion of the mission and reliability. Mission area of a single unmanned vehicle is restricted due to its sensors and communication range [2], while cooperation of multiple agents makes it possible to act in much wider area. Each agent can do the mission in a part of the whole area with respect to its sensor and communication range, and then sharing information with other agents, they can collaboratively accomplish the mission over the whole area in a large-scale application. In addition, there is no need to deploy large-scale UAVs capable of long-endurance flight which means that there is no need to fueled UAVs. So, the cost and radar cross-section will considerably decrease [3]. The speed of mission completion also will increase considerably with higher number of agents [4]. Moreover, when one agent fails for any reason, not only will the whole task not fail but other agents will also complete the mission by reconfiguration in their distribution.

Due to high maneuverability, large-scale operational range, high security and the possibility of accessing hard-to-reach places, UAVs equipped with a computer vision system are very suitable and practical options for monitoring firefighting operations or even automatically executing firefighting missions [5]. With respect to the significant advances made in electronics, computer science and digital cameras, and on the other hand, considering the capabilities mentioned for UAVs, which enable the monitoring of firefighting operations as well as remote control with minimal risk to life, remote sensing has become one of the most common tools in forest monitoring and wildfire disaster management in recent decades [6].

Currently, common remote sensing systems for wildfire monitoring include ground-based monitoring systems such as wireless sensor network or cameras mounted on fixed bases, helicopters and manned aircrafts, as well as satellite imagery [7]. Ground-based systems are not suitable for large-scale areas due to limitations in monitoring range. Satellite images may not have enough resolution to extract details. Helicopters and manned airplanes are also expensive and increase costs. All these factors along with the high capabilities of UAVs that have already been mentioned, work hand in hand to make remote sensing with the help of UAVs equipped with computer vision the best, most effective, least dangerous, safest and cheapest solution for firefighting.

Despite the capabilities that drones can add to the wildfire control team such as the ability to fly to hard-to-reach places, saving time which is very important to reduce damages and casualties, and also send on-line videos from the scene to the crisis control unit and operational teams, compared to the other mentioned methods, suffer from a short life span especially the inexpensive small drones. To solve this deficiency and fix this defect in this amazing solution to low-risk remote firefighting, the use of collaborative multi-agent system can be very attractive and pioneering. Multi-UAV systems can search the area rapidly due to their cooperation, and as a result, they accomplish the initial scanning in the shortest time. They can tell the disaster control unit or operational teams where the most critical points are, where infrastructures suffer the most damage and where the most casualties are. They can also just fly over the area and monitor the surveillance region or gather information and forecast live streaming to the disaster control unit. It may also be necessary to transport or deliver cargo such as food or medicine from point to point. Moreover, UAVs can build a comprehensive map of the area cooperatively which includes important information for relief actions such as the location of victims or survivors and damaged infrastructures. They can track the fire front and predict the fire behavior which can help the operational teams to prevent the fire from spreading further.

For intelligence and autonomous firefighting, the drone must be equipped with sensors so that it can record images and extract information by image processing techniquecs. Digital cameras, infrared cameras or a combination of the two are among the options that are used to analyze the color, motion and geometry of fire. Infrared cameras can be used in situations where the light is low, or there is a lot of smoke. To increase the accuracy, reliability and robustness of fire detection algorithms and decrease the false rate, the combination of these two cameras can be used as well [6].

However, the use of drones in firefighting especially large-scale fire such as forest fire may also come with limitations. In such situations due to large-scale fire, the land surface temperature may reach up to 1300 °C [8] which means that the surrounding temperature reaches several hundred °C and can extremely damage the battery of the UAVs or reduce the battery life and so the endurance of the UAVs [9]. This issue can be solved by using swarm of UAVs as explained before. Due to UAVs high maneuverability, there is no concern about the hardware equipment such as camera which is necessary for image processing techniques in UAVs’ autonomous operations. In [8], a InReC R500 camera is deployed which its temperature range is −4–500 °C. Deploying image processing techniques, the authors detected fire at altitude of more than 15 m among thick smoke. In a large-scale fire area such as forest fire, the surveillance region will be filled with thick and huge smoke. Therefore, using infrared cameras allows the UAVs to see through smoke and monitor the behavior and development of the fire, detect and extract critical sections and also localize or track targets. However, since one of the factors which cause fire spread in large-scale areas such as forests is strong wind, stability of UAVs over fire is a challenging issue.

2 Related works

Forest fire is a natural disaster which usually starts by human negligence or arson and can burn thousands of square kilometers. Firefighting is a very dangerous operation in which the fire front has to be identified to stop the fire through analyzing its behavior and development and then detecting the next probable points. Lack of information about fire front usually leads to the increase in damages and economic losses as well. Moreover, the ability to detect the boundaries of the wildfire is a crucial aspect in a wide area. According to what mentioned in the previous section, using a cooperative multi-UAV system can be an attractive and efficient solution for wildfire fighting. Thus, in [10], the authors presented an autonomous system that works in real time to estimate the propagating boundary. For this purpose, a fleet of low-cost unmanned aerial vehicles that can be deployed quickly were proposed to be used in monitoring a wildfire. The authors in [11] designed a distributed control framework for a collaborative team of multiple UAVs to monitor the wildfire closely and track its development precisely. A changing environment was assumed, and the UAV team was designed in a way that could avoid in-flight collisions and cooperate with each other to achieve the main goal which was maximum coverage over the affected area, tracking the fire front and predicting its development. In [12], the authors proposed deployment of multiple cooperative fixed-wing UAVs for monitoring, behavior analysis and predicting the development of the fire front. A deep learning method was used to control the UAVs motion in order to achieve maximum coverage during forest fires. The authors of [13] proposed a solution for path planning of multiple cooperative fixed-wing UAVs for fire front development monitoring. The terrain, fire propagation process and dynamics of the agents are modeled. A fleet of UAVs equipped with thermal cameras is considered, and agents' flight is assumed at various fixed altitudes. However, in a firefighting mission in regions such as forest or urban environments, there are many obstacles such as trees and tall buildings, and it is necessary for the multi-agent system to be able to avoid these obstacles. In [14], the authors addressed obstacle avoidance for a swarm of UAVs in a search and rescue mission.

In order to successfully accomplish a firefighting mission, the operational team must have a good understanding of the situation in the disaster area and have sufficient information about the environment in order to be able to plan any operation efficiently. Therefore, an auxiliary aerial team consisting of multiple UAVs can provide very useful information during operations by covering the environment and tracking fire front. The authors of [15] have proposed a multi-UAV system for cooperative wildfire coverage and tracking. In [16], authors have also developed a multi-UAV distributed decisional architecture in the framework of the AWARE Project together with a set of tests with real unmanned aerial vehicles (UAVs) and wireless sensor networks (WSNs) to validate this approach in disaster management and civil security applications. There were two groups of UAVs ready to fly. The first one was equipped with an infrared camera, and the second one with the node deployment device was charged with three sensors. Missions performed with real UAVs in the experiments of the AWARE Project were carried out in 2009. Implementation of the mission comprises sensor deployment, fire detection and making fire stop burning and also multi-UAV surveillance. The results of an automated firefighting mission using a heterogeneous team of UAVs are presented in [17] as well. Multiple UAVs with different capabilities were used to collaboratively detect fire on the ground or on the facade of tall buildings and then try to extinguish the fire by throwing a blanket or spraying water.

The aerial aiding system can forecast the gathered information to the disaster control unit for on-line planning and acting based on this information as well. An aerial sensor network presented in [18] consists of small-scale, battery-powered and wirelessly connected UAVs equipped with cameras for disaster management applications. In the proposed multi-agent system, multiple quadcopters fly in the formation to accomplish a predefined mission collaboratively. Cooperative UAVs will fly over a disaster area such as wood fire or a large traffic accident and forecast information to the disaster management team in the form of high-quality sensor data such as images or videos. The work of [19] proposed a multi-agent system composed of multiple quadcopters for wildfire monitoring. The authors considered a cooperation problem between multiple quadcopters as a distributed network that a consensus-based solution was presented for it. In [20], multiple cooperative quadcopters are deployed as a multi-agent system for surveillance and monitoring a wildfire in a large-scale area. Based on physical model of fire front development, a propagation model and also a Kalman filter-based method is presented to estimate the rate of fire development. The problem is solved by a consensus-based method. A multi-agent system comprising multiple cooperative UAVs is presented in [21] for wildfire monitoring in remote and hard-to-reach areas. For this purpose, leader-follower-based distributed formation flight is designed in such a manner that multiple UAVs cover affected area collectively. The leader is a fixed-wing UAV which deploys multiple rotary-wing UAVs which fly at a fixed altitude and over a specific area that is dedicated to them. Each rotary-wing UAV gathers information and captures images from its under-coverage area, and as a result, the whole affected area is covered by the proposed heterogeneous multi-agent system.

However, comparative study with researches introduced in the literature is addressed in Table 1.

Table 1 Comparison of the presented scheme with researches studied in the literature
Full size table

As far as the authors have studied, none of the researches have proposed a heterogenous cooperative multi-UAV system for a two-phase firefighting operation. The main contributions of this paper are summarized as follows:

  1. 1)

    Developing a distributed cooperative coverage control scheme for quickly and initially assessment of the surveillance region by fixed-wing UAVs and extracting critical sections wherein more detailed information gathering is needed.

  2. 2)

    Presenting a distributed cooperative search algorithm using multiple cooperative quadcopters to gather more detailed information in critical sections extracted by fixed-wing UAVs.

  3. 3)

    Proposing a two-phase firefighting operation deploying heterogenous multi-UAV system consisting fixed-wing UAVs and quadcopters.

In summary, in this paper, we address a cooperative coverage control problem deploying a heterogeneous multi-agent system. In order to prevent the spread of fire, it is very important that the operational team can first have a comprehensive map of the amount of damage and casualties so that based on that they can prioritize and plan for action in more critical sections. For this purpose, we considered a heterogeneous team consisting of multiple fixed-wing UAVs and multiple multi-rotor UAVs. First, the team of fixed-wing UAVs fly over the fire area and along with a general scan of the environment, they produce an initial map that includes critical areas based on the density of human casualties and fire severity. Then, the team of multi-rotor UAVs go to the critical areas, fly at a lower altitude and hover if necessary to extract more detailed information.

The remainder of the paper is structured as follows. In Sect. 3, the distributed receding horizon cooperative search and coverage control problem is formulated. A two-phase firefighting scenario is designed in Sect. 4. In the first phase of the mission, a cooperative team of fixed-wing UAVs are deployed for quick scanning of the surveillance region and extracting critical sections. In the second phase, a cooperative team of multiple quadcopters are used for more detailed information gathering since they are capable of hovering at lower altitudes. Finally, Sect. 5 concludes this work.

3 Problem formulation

Solving the problem of collaborative search and coverage control for various applications of multi-agent systems, especially search and rescue, monitoring, surveillance, information gathering and logistics, would be very effective and helpful. The main goal in the problem of cooperative search and coverage control is to program the movement of agents in such a way that while observing different points of the environment and ensuring maximum coverage, they minimize the uncertainty in knowledge of the multi-agent system about the environment and at the same time localize exactly the targets randomly distributed in the environment [22].

In the cooperative search and coverage control problem, at first none of the agents have any knowledge of the environment, and they gradually increase their knowledge by observing different points of the environment. Agents choose their next path in such a way that they go to points that have not been observed by other agents, and consequently have not been discovered yet, or the probability of the presence of the target in those points of the environment is higher than other points. For this purpose, we used two probability maps, TPM and UM, to build a cognitive map in which TPM is a target probability map and UM is uncertainty map. Figure 1 depicts the steps of the distributed cooperative search and coverage control algorithm which includes five steps as follows:

  • Step 1:

Fig. 1
figure 1

Steps of the distributed cooperative coverage control scheme for autonomous search task in forest firefighting

Full size image

Each of the agents builds and updates the TPM based on their current position and their observation of the environment according to [23]:

$${p}_{i,c,k}= \left\{\begin{array}{c}\frac{{p}_{d}{p}_{i,c,k-1}}{{p}_{d}{p}_{i,c,k-1}+{p}_{f}(1-{p}_{i,c,k-1})} , c \epsilon {\mathbb{C}}_{i,k} and {Z}_{i,c,k}=1\\ \frac{(1-{p}_{d}){p}_{i,c,k-1}}{(1-{p}_{d}){p}_{i,c,k-1}+{(1-p}_{f})(1-{p}_{i,c,k-1})} , c \epsilon {\mathbb{C}}_{i,k} and {Z}_{i,c,k}=0\\ {p}_{i,c,k-1 }, c \epsilon \notin {\mathbb{C}}_{i,k}\end{array}\right. (1)$$

In which c is the cell number, and \(Z_{i,c,k} = 1\) indicates that the agent \(A_{i}\) at time step k has observed the target in cell c and also \(Z_{i,c,k} = 0\) means that it believes that there is no target in cell c at time step k according to \(A_{i}\). Furthermore, \({\mathbb{C}}_{i,k}\) is the set of cells covered by agent \(A_{i}\) at moment k which is calculated as follows:

$$ {\mathbb{C}}_{i,k} = \left\{ {c {\Omega } : \begin{array}{*{20}c} {\alpha_{c} - \alpha_{i} \left( k \right)} \\ \end{array} \le R_{s} } \right\} $$
(2)

In which \(\begin{array}{*{20}c} {\alpha_{c} , \alpha_{i} \left( k \right)} \\ \end{array}\) and \(R_{s}\) are coordinate of tell c, coordinate of the agent \(A_{i}\)’s position at time step k and sensing radius of the sensors, respectively, and \({\Omega }\) represents the set of all cells in to which the surveillance region is divided.

Moreover, \(p_{d}\) and \(p_{f}\) the detection and false probabilities which define the sensor performance are as follows:

$$ P{(}Z_{i,c,k} = 1{ | }TE_{c} = 1) = p_{d} \& P{(}Z_{i,c,k} = 1{ | }TE_{c} = 0) = p_{f} $$

In which \({\text{TE}}_{c} \varepsilon\){0, 1} implies the presence or absence of the target in cell c so that \({\text{TE}}_{c} = 1\) and \({\text{TE}}_{c} = 0\) indicate that the target is present and absent in the cell c, respectively.

In addition, based on their current positions and dynamic models, they predict their positions up to the next T time steps. Kinematic movement equations of the agents are considered in the following simple way:

$$ \begin{aligned} x_{i} \left( {k + 1} \right) & = x_{i} \left( k \right) + v_{c} .{\Delta }t.\cos \psi_{i} \left( k \right) \\ y_{i} \left( {k + 1} \right) & = y_{i} \left( k \right) + v_{c} .{\Delta }t.\sin \psi_{i} \left( k \right) \\ \end{aligned} $$
(3)

In which \(x_{i}\) and \(y_{i}\) are x and y components of the agent \(A_{i}\)’s position, \(v_{c}\) is the constant cursing speed, \({\Delta }t\) is time step and \(\psi_{i} \left( k \right)\) is the heading angle of agent \(A_{i}\) at time step k.

The movement of the agents is assumed to be in 2D x-y plane and at a constant altitude and speed, and so the heading angle is the only input to control the movement of the agents. It is also assumed that each agent has only three options for changing the direction, in other words, it can go straight forward without changing the angle, or it can turn clockwise at an angle of 45 degrees and go forward, or it can turn at an angle of 45 degrees counterclockwise and go forward. Therefore, there are three choices in each time step, and if we want to predict the next T time steps, there will be 3T candidate paths for each agent.

At the same time, they send their current position to the neighboring agents and also receive the current position of the neighboring agents.

  • Step 2

Now, based on the shared information, they update the TPM once again:

$$ Q_{i,c,k} = \mathop \sum \limits_{j = 1}^{N} \omega_{i,j,k} {\text{ln}}\left( {\frac{1}{{p_{i,c,k} }} - 1} \right) $$
(4)

\(\omega_{i,j,k} = \frac{1}{{N_{i} }}\) if j \( N_{i} \left( k \right)\) or otherwise \(\omega_{i,j,k} = 0\).

In which \(N_{i} \left( k \right)\) represents the set of agents that are in the communication range with agent \(A_{i}\).

Then, the agents build the UM based on updated TPM:

$$ \eta_{i,c,k} = e^{{ - \vartheta \left| {\begin{array}{*{20}c} {Q_{i,c,k} } \\ \end{array} } \right|}} $$
(5)

In which \(\vartheta\) is a positive constant. On the other hand, each of agents predicts neighboring agents’ positions up to the next T time steps, based on their current position.

  • Step 3

According to its current position and that of its neighbors as well as its next T time steps movement and that of its neighbors, and based on the updated TPM/UM, each agent calculates the optimal path through the following cost function in such a way that while guaranteeing non-collision path, it goes to a point that reduces the uncertainty and increases the probability of identifying the target so that, with the collaboration of other agents, the uncertainty is minimized, and the coverage is maximized finally.

$$ J\left( {i,l,k} \right) = \mathop \sum \limits_{q = 1}^{T} \mathop \sum \limits_{{c \varepsilon {\mathbb{C}}_{i} (p_{i}^{l} (k + q|k))}} \eta_{i,c,k} $$
(6)

s.t. min(dij) \(\ge\) dcollision

In which \(J\left( {i,l,k} \right)\), \(p_{i}^{l} (k + q|k)\) and \(\eta_{i,c,k}\) are cost function related to the l’th candidate path of \(A_{i}\) at time step k, target probability according to \(A_{i}\)’s observation in the l’th candidate path at time step k + q, based on information up to time step k and uncertainty of cell c according to \(A_{i}\) at time step k, respectively. Moreover, dij stands for distance between agents \(A_{i}\) and \(A_{j}\), and dcollision is the safe distance which ensure the collision-free path.

  • Step 4

Each agent acts based on the calculated optimal control input and goes to the new point and registers a new observation.

  • Step 5

The loop is repeated based on the new observation and the new position.

4 Scenario design and simulation results

As explained in Sect. 2, multi-agent systems can be deployed for several various applications. In this paper, we design a two-phase firefighting scenario as depicted schematically in Fig. 2, considering a heterogeneous multi-agent system to accomplish a search and rescue mission in a forest firefighting operation in which a cooperative team of multiple fixed-wing UAVs scan the affected area and determine the most critical sections with the most economic cost and the most victims, and then, separate multiple cooperative teams of quadcopters will be sent to the most critical regions to gather more detailed information and determine the exact location of the victims or survivors and damaged infrastructures as well. All simulations are conducted in Matlab/Simulink on a Laptop equipped with an Intel(R) Core (TM) i7-9750H CPU @ 2.60 GHz, 16 Gb of RAM memory and a GeForce® GTX 1050 with 4 GB GDDR5.

Fig. 2
figure 2

Heterogeneous cooperative multi-agent system for semi-autonomous firefighting; a cooperative team of fixed-wing UAVs for initially rapid scan and a cooperative team of multirotor UAVs for more detailed information

Full size image

The disaster management team can then use this information and send rescue team or a cooperative team of multiple UGVs. As another suggestion, quadcopters can be equipped with supplies such as food or drug and drop them for injured people to survive.

4.1 Phase 1; quickly and initially scanning

In the first phase of the mission, we considered a \(2 km \times 2 km\) surveillance region which has to be explored by a cooperative team of four fixed-wing UAVs. The parameters used for this simulation are listed in Table 1. There are three critical rectangular regions which the cooperative multiple fixed-wing UAVs have to discover for more detailed information gathered by the next cooperative multi-agent systems that comprise quadcopters.

Figure 3 shows the target probability map, TPM, when the global average uncertainty reached 0.1. The probability of the presence of a target in all the cells of the environment. According to each of the agents’, observation at the beginning of the mission is equal to 0.5, and gradually when the agents patrol the environment and observe new cells, they update this number. In this way, if there is a target in that cell, the probability of the presence of the target tends to 1 from that agent's point of view, and if there is no target, it tends to 0, and the agent is sure of the presence or absence of the target in that cell, and that cell is no longer uncertain and will be discovered. Thus, over a period of time, uncertainy of the multi-agent system’s knowledge about the surveillence region will move from 1 to 0 which means that the area is explored, and maximum ceoverage is acheived.

Fig. 3
figure 3

Target probability map; all three critical areas are detected (TPM approaches to 2.1972 and -2.1792 represent the absence and existence of the target in the cell, respectively)

Full size image

The global average uncertainty and the global coverage percent for both cooperative and non-cooperative TPM update methods are depicted in Fig. 4. As described before, in cooperative scheme, UAVs share information with their neighbors. As shown in Fig. 4, in cooperative scheme, the uncertainty and coverage percentage converge to acceptable minimum and maximum values, respectively, much faster than non-cooperative method.

Fig. 4
figure 4

Comparison of cooperative and non-cooperative coverage control schemes

Full size image

Figure 5 shows that the collision avoidance is guaranteed as the minimum distance between the UAVs is always more than 40 m, which means that two UAVs are never in the same cell. The blue line is minimum distance among all agents at each time step, and the red line shows the safe distance; if the distance of the agents is less than this value, non-collision path will be ensured.

Fig. 5
figure 5

Collision avoidance of fixed-wing UAVs during their collaborative operation

Full size image

4.2 Phase 2; more detailed information gathering and cooperative search for targets

When these three critical areas are identified, three separate cooperative multi-agent systems will be sent from the closest UAV stations to gather more detailed information. For each three critical Sects. 1, 2 and 3, we considered four cooperative quadcopters. Sections 1, 2 and 3 are a \(400 {\text{m}} \times 600 {\text{m}}\), \(480 {\text{m}} \times 360 {\text{m}}\) and \(400 {\text{m }} \times 400 {\text{m}}\) rectangular region, in which there are 23, 18 and 10 targets, respectively, that three separate team of four cooperative quadcopters have to extract their exact locations. Simulation will stop when the global average uncertainty satisfies \(\overline{\eta } \le 0.01\). The detail of the paramters used in these three simulations is listed in Tables 2 and 3.

Table 2 Description of parameters used in simulation of initial rapid scan of the forest fire deploying four fixed-wing UAVs
Full size table
Table 3 Description of parameters used in simulation of search task in three critical sections, by a cooperative team of quadcopter UAVs
Full size table

As it is shown in Fig. 6, the global average uncertainty is 0.01 at time t = 932.5 s. The UAVs trajectories, uncertainty map and target probabbilty map at time t = 932.5 s are depiceted in Fig. 7.

Fig. 6
figure 6

Global average uncertainty; critical Sect. 1

Full size image
Fig. 7
figure 7

Cognitive map and UAVs trajectories at time 932.5 s in critical Sect. 1

Full size image

As indicated in Fig. 8, the minimum distance between the UAVs is always more than 10 m which means that the collision avoidance is guaranteed since two UAVs are never in the same cell in the first identified critical section.

Fig. 8
figure 8

Collision avoidance of UAVs during their collaborative operation in critical Sect. 1

Full size image

The global average uncertainty \(\overline{\eta } \le 0.01\) is satisfied at time t = 633.9 as indicated in Fig. 9. The UAVs trajectories, uncertainty map and target probabbilty map at time t = 633.9 s are depiceted in Fig. 10.

Fig. 9
figure 9

Global average uncertainty; critical Sect. 2

Full size image
Fig. 10
figure 10

Cognitive map and UAVs trajectories at time 253.6 s in critical Sect. 2

Full size image

The minimum distance between the UAVs is always more than 10 m as shown in Fig. 11, which guarantees the collision avoidance since two UAVs will never be in the same cell in the second identified critical section.

Fig. 11
figure 11

Collision avoidance of UAVs during their collaborative operation in critical Sect. 2

Full size image

Figure 12 implies that the global average uncertainty aproaches to 0.01, and the maximum ceverage is achieved at time t = 461.5 s.

Fig. 12
figure 12

Global average uncertainty; critical Sect. 3

Full size image

The UAVs trajectories, uncertainty map and target probabbilty map at time t = 461.5 s are depiceted in Fig. 13. Every two UAVs will never be in the same cell in the third identified critical section since the minimum distance between the UAVs is always more than 8 m as it is shown in Fig. 14, which guarantees the collision avoidance.

Fig. 13
figure 13

Cognitive map and UAVs trajectories at time 461.5 s in critical Sect. 3

Full size image
Fig. 14
figure 14

Collision avoidance of UAVs during their collaborative operation in critical Sect. 3

Full size image

5 Conclusion

In this paper, we presented a distributed predictive control scheme for a cooperative team of heterogeneous UAVs to collaboratively explore a disaster area and search for exact location of probable victims or survivors. The first main contribution of this research is developing a distributed predictive coverage control structure for a cooperative heterogeneous multi-agent system in a firefighting disaster scenario. The UAVs communicate and share information with their neighbors and cooperatively build the cognitive maps including TPM and UM. Based on updated cognitive map, each UAV independently decides about its next movement in such a manner that maximize the coverage and minimize the uncertainty without any collisions collaboratively. Each UAV predicts its possible movements few time steps ahead and neighboring agents’ candidate paths as well. Decision structure is such that each agent then solves an optimization problem to choose the optimal path from those candidate paths, based on cognitive map and collision avoidance with neighbors’ possible routes. In this scenario, we certified the efficiency and capacity of the multi-UAV system to be deployed by disaster management team for various applications such as monitoring, search and rescue, information gathering and logistics. The second contribution is using heterogeneous multi-agent system in a two-phase mission, comprising fixed-wing UAVs for generally and initially fast scanning of the environment in the first phase of the mission and rotary-wing UAVs to gather more detailed information and localize exactly the probable victims or survivors in the critical sections prioritized by cooperative fixed-wing UAVs, in the second phase of the mission.

In the future work, we will employ the Monte Carlo simulation to assess the effect of the number of UAVs, the sensing radius, the detection and false alarm probabilities and the communication range on the presented scheme and calculate optimal values of each. We also will extend the cooperative search and coverage algorithm for moving targets and also fault-tolerant control. Currently robustness of a multi-agent system to the agents' failure is so challenging. During or after wildfire disaster, failure of one or more agents is highly probable. So it is very important that a coverage control of a multi-agent system be robust to one or more agents’ failure. In such case, a fault-tolerant control has to be designed, so the rest of the system is able to accomplish the mission by automatic reconfiguration.