A Learning-Based Path Planning Method for Spatial Robots

Space robots face challenges such as communication delays, unknown environmental information, complex terrain and obstacles, and energy constraints when performing path planning. To address the problem of high path costs in sampling-based path planning algorithms due to random sampling in unknown complex scenes, this paper proposes a learning-based path planning method for space robots. This method enables space robots to perform efficient and low-consumption path planning. It constructs a probabilistic network of end-to-end encoder-decoders, using U-net as the thematic framework. The probability distribution predicted by the network serves as the prior information for optimal path planning. Subsequently, the traditional sampling-based path planning algorithm is employed to plan the optimal path based on the probability guidance. Experimental results demonstrate that the method proposed in this paper exhibits superior overall performance in terms of path length, time cost, and memory utilization compared to the traditional RRTs algorithm and the existing CNN + bidirectional search algorithm.