Intention-Oriented Fast Joint Prediction and Planning for Safe Driving

The primary challenges for Autonomous Vehicle (AV) navigation in complex scenarios are to effectively handle the interactions between the Ego-Vehicle (EV) and surrounding participants and generate safe yet non-conservative plans. Current approaches employ Planning-Informed Prediction (PIP) and Joint Prediction and Planning (JPP) to model these interactions. At the same time, information loss exists in PIP and JPP always ignores the intention interactions for a free space generation. To this end, this paper incorporates intention interactions into JPP, where a shared intention interaction module is adopted to both the planning and the prediction tasks via a cross-query mechanism, enabling the model to reduce the model size and accelerate inference speed. In contrast to previous methods that typically plan a single trajectory and result in conservative behaviors, we introduce a contingent-masked planner to balance the safety and efficiency, so as to fulfill the scene-level prediction and greatly improve computational efficiency. We evaluate our approach through closed-loop experiments on the CARLA longest6 and Town05 benchmarks, and faster inference speed and more robust performance against state-of-the-art methods.