Vision-Based Local Motion Planner in Unstructured Environments Using Semantic Segmentation Guided Traversability Modeling and Parallelized Model Predictive Path Integral Algorithm

Vision-based local motion planning in unstructured environments leverages real-time visual data for adaptive pathfinding, enhancing navigation in complex settings. However, current methods face challenges in generalizing traversability estimation in novel scenes and employ a two-stage framework that hinders parallelization and complicates training. This paper introduces a novel vision-based local motion planner using semantic segmentation-guided traversability modeling and a parallelized model predictive path integral (MPPI) algorithm. By transforming the traditional semantic segmentation classification problem into a regression problem with L1 loss and implementing a parallelized MPPI algorithm using the Numba library, generalization is enhanced and the optimal control process is expedited. Results show the proposed method improves motion control success rates by 5% and 29% on the RUGD and RELLIS-3D datasets, respectively, and reduces path length and travel time by 9%–11.4% and 16%–26.7%. The parallelized MPPI algorithm ensures real-time performance, finding the optimal motion control sequence within 0.1 s for the entire trajectory.