Decentralized Coordination of UAV Multi-Agent Systems Using Behavior Trees : A Simulation-Based Study of Auction-Based Task Allocation

Sammanfattning

Teleoperated systems are increasingly being used for tasks that are dangerous, repetitive, or difficult for humans to perform directly. Unmanned aerial vehicles (UAVs) are often used in scenarios where an operator needs to have a bird's-eye perspective and the ability to cover an area quickly, such as in search and rescue, disaster response, or reconnaissance missions. Increased autonomy could reduce the cognitive load on the operator while also making it possible for several UAVs to cooperate with less human control. Coordinating multiple UAVs is challenging, especially in environments with unreliable communication links or dynamically changing conditions, which makes it difficult to rely on a centralized unit to coordinate the mission. To address this, a decentralized behavior tree-based architecture is proposed for a reconnaissance mission in a simulated environment. Each UAV runs its own behavior tree, which integrates search behavior, task allocation, task execution, and coordination with other UAVs. When a task needs to be allocated, the detecting UAV initiates an auction-based task allocation process using a modified version of the Contract Net Protocol, allowing other UAVs to bid on the task. The evaluation considers three aspects of the system: its ability to complete the mission and recover from UAV failures, its scalability as the number of UAVs increases, and the effect of heterogeneous team compositions with different capabilities. The results show that behavior trees can be used to model a decentralized multi-agent system of UAVs, and that the proposed architecture was able to complete the designed mission scenario without relying on a centralized coordinator. The system was also able to recover from UAV failures during a mission, although the failure scenarios increased the mean mission completion time compared to normal operation. Increasing the number of UAVs reduced the mission completion time from 1607 seconds with three UAVs to approximately 1380-1440 seconds with six to eight UAVs, while communication overhead increased with the number of UAVs. The results also show that role specialization alone did not necessarily improve performance, as the outcome depended on how the task allocation strategy took the capabilities of the UAVs into consideration.