Adaptive Formation Tracking of Swarm Jumping Robots Using Multiagent Deep Reinforcement Learning

Swarm jumping robots have attracted increasing attention due to their multimodal motion performance, fault tolerance, and high efficiency. Formation maintenance is the core part of cooperative control of swarm robots, but the mainstream strategies still suffer from poor dynamic adaptability and difficulty in convergence. Here, we propose an adaptive formation tracking system for insect-inspired swarm jumping robots based on Multiagent Deep Reinforcement Learning with Artificial Potential Field (MADRL-APF). It incorporates a hierarchical control architecture for the jumping robots capable of multiple motion modes (jump, turn left, turn right) to autonomously track a moving target. We validated the feasibility of the proposed method in simulated environments with/without obstacles. Multiple jumping robots are able to approach, track and catch a moving target while avoiding obstacles and self-organizing formations during different tasks. Compared with the previous multiagent deep reinforcement learning algorithm, the results show that our method has better convergence and stability. In addition, the method exhibits good scalability and can be scaled to swarm robotic systems with more agents.