Configuration spaces (C-spaces) are an essential component of many robot path-planning algorithms, yet calculating them is a time-consuming task, especially in spaces involving a large number of degrees of freedom (DoF). Here we explore a two-step data-driven approach to C-space approximation: (1) sample (i.e., explicitly calculate) a few configurations; (2) train a machine learning (ML) model on these configurations to predict the collision status of other points in the C-space. We studied multiple factors that impact this approximation process, including model representation, number of DoF (up to 42), collision density, sample size, training set distribution, and desired confidence of predictions. We conclude that XGBoost offers a significant time improvement over other methods, while maintaining low error rates, even in C-Spaces with over 14 DoF.

翻译：配置空间（C-space）是众多机器人路径规划算法的重要组成部分，但其计算过程耗时巨大，尤其是在涉及大量自由度（DoF）的空间中。本文探索了一种基于两步数据驱动的C-空间近似方法：（1）采样（即显式计算）少量配置；（2）基于这些配置训练机器学习（ML）模型，以预测C-空间中其他点的碰撞状态。我们研究了影响该近似过程的多个因素，包括模型表示、自由度数量（高达42个）、碰撞密度、样本量、训练集分布以及期望的预测置信度。结论表明，即使在超过14自由度的C-空间中，XGBoost相比其他方法在保持低错误率的同时，能够显著提升计算效率。