ChainFlow-VLA: Causal Flow Planning with Vision-Language Models

Researchers introduced ChainFlow-VLA, a new autonomous driving planner that combines autoregressive causal trajectory generation with diffusion-based refinement.

The method uses vision-language model features to correct and globally optimize trajectories after step-by-step mode prediction.

ChainFlow-VLA reached a top NAVSIM v1 score of 94.85, approaching human-level performance.

The approach targets a major gap in self-driving systems: balancing causal reasoning with consistent, safety-critical trajectory planning.