WCog-VLA: Bridging Foresight for Proactive Autonomy

Existing Vision-Language-Action (VLA) models for end-to-end autonomous driving have been inherently limited by either incomplete world cognition or fragmented foresight, confining them to reactive responses. This fundamental constraint prevents truly intelligent, anticipatory driving behavior.

Dual-Level World Cognition for Proactive Driving

The WCog-VLA framework introduces a novel dual-level approach to overcome the reactive limitations of current VLA models, enabling WCog-VLA proactive autonomous driving. At its core, WCog-VLA bridges semantic world forecasting with generative world evolution. The semantic level unifies world cognition and reasoning by integrating 3D spatial perception and injecting agent tokens to capture dynamic world interactions. This is further enhanced by Game-theoretic Chain-of-Thought (Game-CoT) reasoning, allowing for more strategic decision-making.

Accelerated Generative World Modeling

A critical innovation in WCog-VLA is the Aligned Decoupled Diffusion Transformer (ADDT) at the generative level. This powerful generative world model synthesizes physically-plausible joint multi-agent trajectories. Crucially, ADDT accelerates inference by significantly reducing the required denoising steps through scene representation alignment, addressing a common bottleneck in diffusion models. This efficiency gain is vital for real-time applications like autonomous driving.

Strategic Reasoning and Data Enrichment

To facilitate the advanced strategic reasoning capabilities required for WCog-VLA proactive autonomous driving, the researchers constructed a substantial dataset featuring 85k Game-CoT annotations. This large-scale dataset is instrumental in training models to understand complex multi-agent interactions and anticipate future scenarios. The efficacy of WCog-VLA is underscored by its State-Of-The-Art (SOTA) PDMS score of 92.9 on the NAVSIM benchmark, demonstrating a significant leap in performance for autonomous driving systems.