DriveLaW: Unifying Planning and Video Generation in a Latent Driving World

摘要

World models have become crucial for autonomous driving, as they learn how scenarios evolve over time to address the long-tail challenges of the real world. However, current approaches relegate world models to limited roles: they operate within ostensibly unified architectures that still keep world prediction and motion planning as decoupled processes. To bridge this gap, we propose DriveLaW, a novel paradigm that unifies video generation and motion planning. By directly injecting the latent representation from its video generator into the planner, DriveLaW ensures inherent consistency between high-fidelity future generation and reliable trajectory planning. Specifically, DriveLaW consists of two core components: DriveLaW-Video, our powerful world model that generates high-fidelity forecasting with expressive latent representations, and DriveLaW-Act, a diffusion planner that generates consistent and reliable trajectories from the latent of DriveLaW-Video, with both components optimized by a three-stage progressive training strategy. New state-of-the-art results across both tasks demonstrate the power of our unified paradigm. DriveLaW not only significantly advances video prediction, surpassing the previous best-performing work by 33.3% in FID and 1.8% in FVD, but also sets a new record on the NAVSIM planning benchmark.