PhyDrive: Planning-Oriented Physical Evidence Interaction for End-to-End Autonomous Driving

Planning in end-to-end autonomous driving requires anticipating the future evolution of scenes, motivating the development of world models for informed decision-making. Existing methods primarily exploit latent representations from world models to improve planning, but often lack explicit physical evidence for reasoning about the physical consequences of candidate motions, limiting their ability to assess safety risks. To address this limitation, we propose **PhyDrive**, an end-to-end autonomous driving framework. First, we design a Physical-Evidence-Guided Scorer (PEGS) that extracts planning-oriented physical evidence and uses it to guide interactions between candidate-specific latent representations and the global scene context, enabling safety-aware trajectory scoring and selection. Second, we introduce a Planning-oriented Latent Representation Learning (PLRL) strategy with an training-only auxiliary network that encourages the world model to encode planning-relevant distinctions, thereby improving its sensitivity to motion-specific scene interactions. On the NAVSIM v1 and v2 benchmarks, **PhyDrive** achieves state-of-the-art performance, with PDMS and EPDMS scores of 95.3 and 93.0, respectively. The code will be publicly available.