Reinforcement learning-based reliable path planning for off-road autonomous ground vehicles in unknown, uncertain environments

Path planning for Autonomous Ground Vehicles (AGVs) in off-road environments is challenging due to the inherent uncertainty in terrain traversability and obstacle presence. Traditional deterministic methods often fail to guarantee mission reliability under such unknown and uncertain environments. This article proposes an uncertainty-aware Reinforcement Learning (RL) path planning framework based on local sensing to achieve reliable path planning in unknown and uncertain environments under a probabilistic-obstacle abstraction of terrain traversability risk. The uncertain environment model is first established by abstracting terrain mobility analysis into a set of randomly distributed probabilistic obstacles. Subsequently, based on the user-specified reliability threshold and local sensing results, a beam-wise probabilistic safety margin is derived to quantify the risk of specific navigation directions relative to the required task success probability. To address the intractable path planning problem with probabilistic obstacles, a Monte Carlo sampling-based method is employed to evaluate the path success rate. Finally, a Soft Actor–Critic (SAC) RL agent is designed, featuring a modular network architecture with independent feature extractors to process structured observations that integrate vehicle dynamics, goal information and the proposed probabilistic safety metrics. Furthermore, a comprehensive reward function incorporating success probability constraints is designed to balance path efficiency and reliability. Simulation results in a 2D environment demonstrate that the proposed framework effectively generates paths that satisfy varying success-probability requirements and outperforms deterministic baselines in the simulated uncertain scenarios.