Heuristic Sampling for Robot Motion Planning via Conditional Generative Adversarial Networks in Narrow Spaces
Xingchen Li, Xifeng Gao · 2024
Sampling-based motion planning methods are widely used in the robotics community but often exhibit diminished performance in narrow spaces. Despite efforts to enhance planning efficiency by introducing various sampling strategies aimed at increasing the likelihood of sampling within narrow spaces, these approaches are often tailored to specific planning problems, limiting their generalizability. Here, we present a motion planning algorithm employing heuristic sampling via Conditional Generative Adversarial Networks (CGAN), which leverages insights from previous successful plans to predict potential regions containing feasible paths for new planning problems. By utilizing such prior information to guide sampling, our approach enables motion planners to rapidly identify feasible paths, even in complex environments. Through extensive simulation experiments, we demonstrate the computational efficiency of this sampling strategy.