A Reinforcement Learning based RRT Algorithm with Value Estimation

摘要

Path planning algorithm is always a heated area of robotics. Researchers propose various of algorithms to meet different requirements. Yet a remaining issue for path planning in unknown environment is that little information is available. For these occasions, researchers proposed sampling-based path planning algorithm such as Rapidly-Exploring Random Trees. This kind of method relies greatly on the sampled points. Noticing that reinforcement learning methods learn about the environment during the interaction process, it is manageable to combine these two methods to improve algorithm’s behavior. In this manuscript, a reinforcement learning based RRT algorithm is proposed to search path in environment with little previous information. The proposed method uses value estimation from reinforcement learning to encourage exploration and makes the agent sample points from less visited area. According to simulation results, the proposed algorithm has higher utilization of the tree nodes and explores more area comparing with RRT algorithm.