VFlash Cuts Robot Visual Planning Latency 25x via Diffusion

Robot policies built on vision-language models spend most of their compute budget spelling out spatial coordinates one token at a time. VFlash, released November 6, 2026, eliminates that bottleneck by substituting a 537-million-parameter diffusion drafter that predicts entire visual primitives (trajectories, bounding boxes, 3D motion plans) in a single forward pass.

The method, called Flash Primitive Decoding, exploits the known geometric structure of spatial outputs to allocate output slots before decoding begins. Because manipulation accuracy does not require reproducing an autoregressive model's exact token sequence, the drafter trains directly against annotated coordinates and accepts any structurally valid prediction without forcing token-for-token alignment with the base model.

Speed and Accuracy Gains

The headline performance claim is approximate: single-rollout 3D point motion forecasting runs roughly 25 times faster, while bounding box decoding shows 2.6 to 2.8 times speedup. On accuracy, a world action model conditioned on VFlash plans achieves 70.2 percent average success across four MolmoSpaces task suites, compared with 62.6 percent for the autoregressive baseline. On InternA1 3D robot trajectories, average displacement error drops 28 to 61 percent under joint 2D and 3D decoding.

The measurements reflect single-rollout tests. The paper does not report what the speedups look like under batched, multi-step rollouts typical of real manipulation loops, leaving open questions about wall-clock performance in production settings.

Flash Primitive Decoding Architecture

Traditional vision-language models for robotics emit coordinates as sequences of discrete tokens, one at a time. VFlash instead trains a 537-million-parameter drafter that takes the vision-language model's internal state and outputs an entire primitive in parallel. The drafter uses a diffusion process to refine its prediction over multiple denoising steps, but all denoising happens within a single forward pass rather than across multiple autoregressive decode cycles.

Because the method decouples spatial precision from token fidelity, it can accept any coordinate tuple that satisfies structural constraints (for example, a bounding box with x-min less than x-max). This tolerance for variation lets the drafter optimize directly for task success rather than sequence likelihood.

Benchmark Results

On the MolmoSpaces manipulation benchmark, VFlash-conditioned policies reached 70.2 percent average success, a 7.6-percentage-point improvement over the 62.6 percent baseline. On InternA1 and AgiBot 3D robot trajectory prediction, average displacement error fell between 28 and 61 percent depending on the task and whether 2D and 3D outputs were decoded jointly.

The work arrives during a sustained wave of spatial-reasoning robot policy research. Emergent's robotics tracker has logged 190 stories in the last 90 days, reflecting growing academic and industry focus on vision-language models that can plan in three dimensions.

Release Date and Availability

VFlash was officially released on November 6, 2026, with the research paper and implementation posted to Hugging Face. The authors do not specify licensing terms or whether pretrained drafter weights are publicly available. The paper cites preliminary single-rollout benchmarks, leaving batched inference performance and real-world deployment results for future work.

What This Means

VFlash demonstrates that autoregressive token decoding is not a hard requirement for spatial reasoning in vision-language robot policies. By predicting entire primitives in parallel, the method trades off exact token reproduction for task-oriented accuracy and wall-clock speed. If the single-rollout speedups generalize to batched, multi-step loops, the approach could reduce the compute overhead of deploying large vision-language models on hardware-constrained robots. The next test will be whether the 25x claim holds under the real-time, multi-object manipulation scenarios that define production robotics workloads.