VFlash decodes robot actions in one shot: 5.1x faster inference, 61% lower 3D trajectory error
DrJimFan · x · 2026-11-06
Jim Fan introduces VFlash, a "System One vision model" for Physical AI. Key idea: embodied prediction outputs (bounding boxes, robot trajectories, dense 3D motion) are fixed-length float vectors — a box is always 6 tokens, a 10-waypoint trajectory 54 — so the whole geometry can be output in parallel instead of decoded autoregressively.
Results:
- Up to 12x faster action decoding, 5.1x overall robot inference speedup on edge GPUs (Jetson Thor, DGX Spark)
- Up to 61% lower 3D trajectory error than the VLM backbone
- The gain comes from avoiding cascading autoregressive errors: "every coordinate is written in ink" with no backspace
Paper and code are open source. His thesis: everything autonomous will need to think on its own chip — physical reasoning at the speed of reflex.
Related event: NVIDIA Releases VFlash, Speeding Up Physical AI Inference 5.1x(4 posts)→
More from Embodied
- RL post-training can improve generalist robot policies on novel unseen tasks — arkrause · 2026-11-06
- Nemotron 3.5 Bolt E4B hits 27.7 tok/s on Jetson Orin Nano Super, 1.8x faster than baseline — Any_Frame9721 · 2026-11-06
- Agent builds a working satellite TV receiver on FPGA with $500 in tokens — appenz · 2026-11-06
- OpenTac brings HDR tactile sensing to solve the range-sensitivity tradeoff — _krishna_murthy · 2026-11-06
- Zipline raises $1.5B at a $21.5B valuation for autonomous drone delivery — lukas_m_ziegler · 2026-11-06
- Developer ships Drivr Glass: voice-first Grok cabin bot now displays cards in-car — bfrench · 2026-11-06