Sampling-Based Path Planning Driven by Transformers with Direct Sampling in Ellipse and Region
Huang Yuan, Tianyu Shen, Cheng-Tien Tsao, Hee‐Hyol Lee · 2023
Path planning problem aims at finding a collision-free path between a starting vertex and a goal vertex. Recently, neural-network-driven path planning algorithms have been developed to predict a region containing an optimal path, which subsequently serves as a direct sampling domain for sampling-based path planners. However, low-accuracy predictions hinder the further development of neural-network-driven algorithms, leading to long computation time and inefficient sampling with a uniform sampler. In this paper, we first design a transformer-based model, namely improved segmentation transformer (ISETR), involving a segmentation task and a regression task. A region containing an optimal path and an estimation of the path length are processed separately. The outputs of the ISETR are employed to realize a direct sampling for subsequent path planning, which is composed of regional and elliptical domains. Distinct from other neural-network-driven algorithms, the elliptical domains derived from the estimated length replace the uniform sampler, guaranteeing a probabilistic completeness of the proposed sampling-based path planning algorithm, ISETR-RRT*. Ablation studies verify the superiority of the ISETR in the accuracy of predictions. Additionally, the ISETR-RRT* outperforms other transformer-based path planning methods and an informed RRT*, in terms of computation time, path length, and sampling numbers.