VFlash: Fast, Accurate Visual Planning via Flash Primitive Decoding

Researchers introduced VFlash, a method for fast, accurate visual planning that decodes points, bounding boxes and trajectories as "flash primitives" rather than generating visual plans one coordinate at a time. The approach targets physical AI tasks such as robotics and autonomous driving, where autoregressive vision-language models are slow because they emit each plan element sequentially. VFlash addresses that bottleneck by decoding primitives in parallel, according to the work.

Models for physical AI tasks, such as robotics and autonomous driving, need not reason solely in language. They can use visual plans--points, bounding boxes, and trajectories--to represent spatial goals and anticipate motion. However, autoregressive vision-language models generate these plans one co