A Learning-based Multi-RRT Approach for Robot Path Planning in Narrow Passages

As an important class of sampling-based path planning methods, the Rapidly-exploring Random Trees (RRT) algorithm has been widely studied and applied in the literature. In RRT, how to select a tree to extend or connect is a critical factor, which will greatly influence the efficiency of path planning. In this paper, a novel learning-based multi-RRTs (LM-RRT) approach is proposed for robot path planning in narrow passages. The LM-RRT approach models the tree selection process as a multi-armed bandit problem and uses a reinforcement learning algorithm that learns action values and selects actions with an improved epsilon-greedy strategy (epsilon (t) -greedy). Compared with previous RRT algorithms, LM-RRT can not only enhance the local space exploration ability of each tree, but also guarantee the efficiency of global path planning. The probabilistic completeness and combinatory optimality of LM-RRT are proved based on the geometric characteristics of the configuration space. Simulation and experimental results show the effectiveness of the proposed LM-RRT approach in single-query path planning problems with narrow passages.