PVSR-IQL: Offline Reinforcement Learning With Predictive Vehicle State Representation for Off-road Path Tracking

This paper proposes PVSR-IQL, an offline reinforcement-learning framework for off-road path tracking under partial observability. Predictive Vehicle State Representation (PVSR) extracts control-relevant vehicle states and a latent representation from histories of observations using state estimation and next-observation prediction. The actor uses measurable observations and PVSR outputs, while the Q and value networks additionally use privileged vehicle and path states during training. This asymmetric structure enables the policy to operate without privileged information during deployment. Closed-loop evaluations were conducted with a mining truck in the CARLA Mine_01 off-road environment. Compared with the baseline IQL method, PVSR-IQL reduced mean lateral error by approximately 34.4% under nominal conditions, 18.7% under initial pose perturbations, and 20.0% and 32.9% under velocity and acceleration observation noise, respectively. For a complex route at an out-of-distribution reference speed of 7 m/s, it reduced the mean lateral error by 52.8%. These results indicate that PVSR improves the lateral path-tracking accuracy and robustness of offline reinforcement learning under varying operating conditions.