BrainWAM's action-space coordination tops NAVSIM leaderboards with 89.6 EPDMS

Autonomous driving demands planning that respects both semantic constraints—like traffic rules and scene understanding—and predictive dynamics, such as how other vehicles will move. Existing end-to-end approaches typically excel at only one side: Vision-Language-Action (VLA) models leverage vision-language priors for reasoning, while World Action Models (WAMs) generate future-aware predictions. A team led by Bing Zhan (including Zhaoxiang Zhang) found that naively combining these through token-level attention causes an attention-allocation mismatch: semantic shortcuts dominate and suppress predictive dynamics.

To solve this, the team introduces BrainWAM, a structured action-space coordination framework inspired by neuroscience. It routes semantic reasoning and predictive world modeling into two specialized action-oriented pathways, aligning them at the level of compact action representations rather than raw attention. An asynchronous rectified-flow inference strategy with decoupled video and action denoising shortens latency while preserving planning-relevant context. BrainWAM reaches state-of-the-art performance on NAVSIM v1 (89.5 PDMS) and NAVSIM v2 (89.6 EPDMS), outperforming both VLA-only and WAM-only methods, making it a practical new direction for autonomous driving systems.