Integrating Driving-Aware World Model With MPC for Autonomous Driving at Unsignalized T-Intersections

Model Predictive Control (MPC) is widely used in autonomous driving for its robustness and control effectiveness. However, its reliance on explicit physical models can lead to mismatches in interactive scenarios, resulting in trajectory deviations and safety risks. While recent world model approaches alleviate this limitation by learning latent dynamics from data, their learned representations often retain task-irrelevant background features, leading to instability propagation in downstream control policies. To address these challenges, we propose an end-to-end framework that integrates a Driving-Aware World Model (DAWM) with Model Predictive Path Integral (MPPI) control. DAWM employs a spatial prior and a scoring network to adaptively reweight Vision Transformer (ViT) patch embeddings, producing compact and task-relevant latent states. It further introduces an incremental latent dynamics model to predict state transitions, and an uncertainty-aware loss weighting mechanism to balance multi-objective training. Meanwhile, MPPI sampling is guided by a reinforcement learning policy and augmented with artificial potential fields (APFs) to enhance safety and convergence. In unsignalized T-intersection scenarios in CARLA, our approach outperforms standard MPPI by 66.7% in task success rate, while also reducing collisions and lane deviation, highlighting its superior control performance.