NVIDIA VFlash speeds robot visual planning up to 12× and cuts trajectory error up to 61%
Interesting Engineering, covering NVIDIA Research's VFlash paper, says parallel flash primitive decoding cut Jetson Thor manipulation latency about 80% versus autoregressive baselines.
Why it matters
VFlash is a decoding framework that lets vision-language models emit whole visual primitives—boxes, 2D/3D trajectories, dense motion—in one parallel draft instead of token-by-token coordinates. NVIDIA reports up to 12× faster robot-planning decoding and up to 61% lower 3D waypoint error versus the autoregressive target on some suites, with roughly 80% lower end-to-end manipulation latency on Jetson Thor and DGX Spark. Interesting Engineering timed its write-up at 7:13 p.m. EST on Nov. 5 (00:13 UTC Nov. 6). Downstream world-action models conditioned on VFlash plans reached 70.2% average success on MolmoSpaces versus 62.6% baseline. This is research, not a product SKU—watch for open weights, TensorRT-Edge releases and independent replications.
Key facts
- Interesting Engineering timestamp Nov. 5, 2026, 7:13 p.m. EST (00:13 UTC Nov. 6)
- Up to 12× faster robot-planning decoding vs autoregressive
- Up to 61% lower 3D trajectory error on reported suites
- ~80% lower manipulation latency on Jetson Thor / DGX Spark
- MolmoSpaces avg success 70.2% with VFlash plans vs 62.6% baseline
- Primary technical page: NVIDIA Research VFlash