Formation Control of Multi-agent Based on Deep Reinforcement Learning

This paper is based on the multi-agent deep deterministic policy gradient (MADDPG) deep reinforcement learning algorithm, combined with the Leader-Follow method to complete the multi-agent circular formation control problem, considering the distance constraints and angle constraints between the agents. Overcome the problem that it is difficult to accurately model objects in previous control methods, and do not need to care about network topology, system order and other preconditions. In addition, for the formation movement problem, we predefine a virtual leader that moves according to a random curve trajectory, and adopt a two-stage training method. Based on the circular formation behavior strategy, each agent continues to train to follow the virtual leader. Simulation experiments verify the effectiveness of the algorithm.