Genetic Fuzzy Methodology for Decentralized Multi-UAV Cooperation with Physical Constraints
Anoop Sathyan, Ou Ma · AIAA SCITECH 2023 Forum · 2023
View Video Presentation: https://doi.org/10.2514/6.2023-1073.vid In this work, we train controllers in a reinforcement learning manner using an evolution based strategy called Genetic Fuzzy Methodology for application to decentralized control of a team of cooperative UAVs to transport a shared slung payload. The objective is to transport a payload to any given target location and maintain it at that location in the 3D space. The controllers on the agents, modeled as fuzzy systems, learn using feedback received from the simulation environment. Genetic Algorithm is used to tune the parameters of the controller to achieve the overall goal of bringing the payload to the desired locations while satisfying the physical and operational constraints of the system. As the system is decentralized and the decisions are made at the agent (UAV) level, the UAVs do not have explicit knowledge of their teammates' actions. Since the training is done in a centralized manner with the fitness values calculated cumulatively for the whole team rather than for individual agents, the agents are able to learn strategies that do not require explicit communication of states and actions between agents. This framework necessitates the individual agents to learn cooperative behavior so that the team goal can be achieved. Once the team is trained, the agents can be tested on new scenarios defined by changing the target locations for the payload. We evaluate performance metrics for the trained system by testing on new scenarios.