Expert-guided Path Planning Based on RRT-star and Deep Reinforcement Learning

The rapidly-exploration random tree star (RRT-star) is widely used in path planning tasks due to its high exploration efficiency. Nevertheless, due to the feature of random sampling in RRT-star, previous experience can hardly be leveraged in subsequent exploration, making it difficult to continuously optimize the path length. Deep reinforcement learning (DRL) can be trained to approach the optimal path under the guidance of a suitable reward. Reward is essential during the training process of DRL. To obtain optimized path planning results, DRL requires suitable rewards. In order to solve this problem, a structure that uses the tunnel reward and combines the advantages of RRT-star and DRL (TRD) is proposed. This structure uses the expert path provided by RRT-star to guide the exploration of DRL and the tunnel reward to encourage the agent to move along the tunnel, guiding the agent to bypass obstacles and continue to shorten the path length. The experimental results have verified the effectiveness and the optimization performance for path length of this structure.