Proprioception-Guided Framework for Terrain Roughness Assessment and Bayesian RRG Planning

Irregular and unpredictable terrains in unstructured environments, including mountains, deserts, rural areas and construction sites, remain a fundamental obstacle to reliable autonomous navigation. The primary difficulty lies in enabling vehicles to accurately understand the risks posed by complex terrains and traverse along flat paths. To address these challenges, this paper presents a terrain-aware path planning framework that couples proprioceptive-sensitive roughness assessment with an improved rapidly-exploring random graph (RRG). From first principles, we design a multi-scale frequency-domain analysis network (MFDAN) that learns terrain roughness from vehicle proprioceptive responses, isolating the frequency bands most responsible for jolts. These bands underpin a terrain roughness map aligned with the vehicle’s dynamic response to the terrain. Building on this map, we introduce a Bayesian learning RRG planner that integrates roughness gradients for high-yield expansion and online Bayesian updating to bias exploration toward safer and more promising directions. The proposed framework is extensively validated in diverse Gazebo simulations and against state-of-the-art baselines, demonstrating significant reductions in jolt level, planning failures, and computation time. Finally, real-world deployment on a 40-ton vibroseis truck in unstructured field environments confirms the practicality and robustness of the framework.