Cumulative Training and Transfer Learning for Multi-Robots Collision-Free Navigation Problems

Recently, the characteristics of robot autonomy, decentralized control, collective decision-making ability, high fault tolerance, etc. have significantly increased the applications of swarm robotics in targeted material delivery, precision farming, surveillance, defense and many other areas. In these multi-agent systems, safe collision avoidance is one of the most fundamental and important problems. Difference approaches, especially reinforcement learning, have been applied to solve this problem. This paper introduces a new cumulative learning approach which comprises of application of transfer learning with distributed multi-agent reinforcement learning techniques to solve collision-free navigation for swarm robotics. In our method, throughout the learning processes from the least complexity scenario to the most complex one, multiple agents can improve the shared policy through parameter sharing, reward shaping and multi-round multi-steps learning. We have adapted two policy gradient algorithms (TRPO and PPO) as the core of our distributed multiagent reinforcement learning method. The performance has shown that our new methodology can help reduce the training time and generate a robust navigation plan that can easily be generalized to complex in-door scenarios.