This is a Plain English Papers summary of a research paper called VFlash Decodes Whole Robot Trajectories in One Pass, Improving Accuracy and Speed. If you like these kinds of analyses, you can find more AI and machine-learning research on AIModels.fyi or follow us on Twitter.
VFlash turns whole spatial plans into one decoding step—and improves their accuracy
VFlash replaces token-by-token coordinate generation with a lightweight drafter that predicts an entire structured visual primitive—such as a trajectory or bounding box—in one pass. Across the evaluated robot trajectory tasks, it reduces errors versus its autoregressive (AR) target; dense 3D point-motion forecasting is about 25× faster in preliminary single-rollout measurements. In downstream MolmoSpaces manipulation, a world action model conditioned on VFlash plans reaches 70.2% average success across four task suites, up from 62.6% for the baseline.
The primitive’s known layout lets VFlash skip token-by-token agreement
A visual primitive has a known schema: its type, coordinate dimensions, and extent determine what each output position means. VFlash uses that schema to build a prediction block for the complete primitive, with the requested length setting the block size for variable-length trajectories. A lightweight diffusion drafter, conditioned on features from a frozen vision-language model (VLM), predicts all coordinate positions jointly; constraints then enforce legal coordinate and delimiter formats....
