WCog-VLA introduces dual-level world-cognitive VLA for proactive autonomous driving

Researchers propose WCog-VLA, a novel dual-level World-Cognitive Vision-Language-Action framework designed to enable proactive autonomous driving by bridging semantic forecasting with generative evolution. The model addresses limitations in existing methods that lack comprehensive world cognition or suffer from fragmented foresight.

At the semantic level, WCog-VLA unifies world cognition and reasoning using 3D spatial perception, agent tokens for dynamics capture, and Game-theoretic Chain-of-Thought (Game-CoT) reasoning. At the generative level, it employs an Aligned Decoupled Diffusion Transformer (ADDT) to synthesize physically-plausible joint multi-agent trajectories while reducing denoising steps to accelerate inference. The authors also constructed a large-scale dataset featuring 85k Game-CoT annotations to facilitate strategic reasoning.

Extensive experiments on the NAVSIM benchmark demonstrate that WCog-VLA achieves a State-Of-The-Art (SOTA) PDMS score of 92.9.