GRRL: Goal-Guided Risk-Inspired Reinforcement Learning for Efficient Autonomous Driving in Off-Road Environment

When unmanned ground vehicles (UGVs) are deployed for rescue, pathfinding, and other tasks in off-road scenarios, existing conventional path planners often fail to generate both feasible and efficient trajectories and also require high computational resources. To address these problems, this article introduces a goal-guided, risk-inspired reinforcement learning framework for efficient autonomous driving in off-road environments. This tightly coupled navigation framework integrates a perception network based on a Transformer feature extractor with an actor-critic reinforcement learning network. First, terrain risk maps are constructed from point clouds and converted into bird’s-eye view risk images as the input of the learning framework. Then, terrain risk features and goal information are combined to generate safe actions for the UGV to move to the goal in off-road environments. Moreover, a joint reward function consisting of goal heuristic and driving feasibility is also designed to ensure both the efficiency and safety of the driving process. Sufficient experiments were carried out on simulated off-road terrains to verify the effectiveness of the proposed system. Experiment results show that, compared with other methods, our method achieves a higher success rate and enhances the UGV’s risk avoidance ability in off-road scenarios.