FutureSightDrive: Thinking Visually with Spatio-Temporal CoT for Autonomous Driving

FSDrive represents a significant advancement in the field of autonomous driving by allowing Vision-Language-Action (VLA) models to process visual information more effectively. Traditional reasoning methods often create a gap between perception and planning, but FSDrive introduces a visual spatio-temporal Chain-of-Thought (CoT) that captures both spatial structures and temporal changes in a unified framework. By functioning as a world model, FSDrive generates future frames that include predicted backgrounds and physically plausible elements, enhancing the model's ability to plan trajectories based on real-time observations. Evaluations on datasets like nuScenes and NAVSIM have shown that FSDrive not only improves trajectory accuracy but also significantly reduces the likelihood of collisions, marking a crucial step forward in the development of safer autonomous driving technologies.