VFlash Decodes Robot Visual Plans ~25x Faster on 3D Motion

Robot policies built on vision-language models spend most of their decode time spelling out coordinates one token at a time. VFlash, posted to Hugging Face on November 6, swaps that habit for a 537M-parameter "drafter" that predicts a whole visual primitive, whether a point trajectory, a bounding box, or a 3D motion plan, in a single forward pass.

The method, called Flash Primitive Decoding, exploits what the authors describe as "the known structure of visual primitives" to fit the output slots before decoding begins. Because spatial accuracy "need not require reproducing an autoregressive model's exact token sequence," the drafter trains against annotated coordinates directly and accepts structurally valid predictions without needing to match the base model's token-for-token output.

The headline speed claim is preliminary. The paper reports "approximately 25× faster generation for 3D point motion forecasting" from single-rollout measurements, alongside roughly 2.6 to 2.8× on bounding box decoding. On accuracy, a world action model conditioned on VFlash plans "achieves 70.2% average success across four MolmoSpaces task suites, compared with 62.6% for the baseline," and on InternA1 3D robot trajectories, average displacement error drops 28 to 61% under joint 2D+3D decoding.

The work lands amid a steady run of spatial-reasoning robot policy research on our robotics tracker, 190 stories logged in the last 90 days. The paper does not report what the one-rollout speedups look like under the batched, multi-step rollouts of a real manipulation loop.