VFlash speeds up visual planning for robot and driving models

Vision-language models used for robotics and autonomous driving can express plans as points, boxes and trajectories, but they emit coordinates one token at a time, so longer predictions get expensive. VFlash, described in a new paper, replaces that loop with Flash Primitive Decoding: a lightweight drafter produces a whole primitive in one forward pass, with a layout that adapts to the primitive's type, dimensionality and length.

It trains on annotated primitives and accepts structurally valid outputs rather than requiring the exact token sequence of an autoregressive model. On 2D and 3D robot trajectory tasks the authors report lower prediction errors, preliminary single-rollout measurements of about 25x faster generation for 3D point motion forecasting, and 70.2% average success for a world action model using VFlash plans across four MolmoSpaces suites, versus 62.6% for the baseline.