HUP: A Holistic Uncertainty-Aware Motion Planner for Autonomous Driving

Uncertainties arising from inherent observation randomness and model limitations significantly affect the robustness and adaptability of motion planning for autonomous vehicles. Existing methods typically incorporate uncertainty into planning outcomes or risk measures but do not explicitly model and propagate it throughout the planning pipeline, leaving its intrinsic influence underexplored. To address this limitation, we propose a holistic uncertainty-aware motion planner (HUP) that explicitly models and propagates uncertainties across the entire pipeline. In the scene representation stage, spatiotemporal positional uncertainties of both static map elements and dynamic traffic participants are uniformly modeled using learnable Laplace distributions. To model interactions among heterogeneous scene elements, we introduce the probabilistic interaction attention module (PIAM), which represents interactions via probabilistic attention, with attention weights sampled from a mixed variational prior comprising two log-normal distributions. Finally, we design uncertainty-aware constraints by formulating collisions and boundary violations as probabilistic events, thereby leveraging scene uncertainties to regulate planned motions. On nuScenes, HUP achieves a highly competitive average $\ell _{2}$ displacement error (DE) of 0.35 m and a collision rate (CR) of 0.15%. In closed-loop evaluation on NeuroNCAP, it achieves the best overall performance among all compared methods, with an average score of 3.52 and a CR of 44.3%. Beyond planning, HUP also improves intermediate perception performance and converges significantly faster than the baseline, benefiting from the complementary regularization induced by jointly optimizing deterministic and uncertainty-aware losses. Moreover, statistical analyses and visualization results show that HUP effectively captures real-world uncertainties in both scene representation and stochastic interaction modeling.