VFlash

VFlash is a System-1 vision-language decoding framework for Physical AI announced on 2026-11-05 by NVIDIA's Learning and Perception Research (LPR) team with GEAR. Project page: https://research.nvidia.com/labs/lpr/vflash/ . Jim Fan (@DrJimFan) and Zhiding Yu posted the launch the same day.

Insight: robots and AVs must continually refresh visual plans (grasp points, trajectories, scene evolution), but autoregressive coordinate-token generation gets costly. VFlash exploits the known geometry of visual primitives so a light drafter can emit a whole primitive in one parallel pass, while a frozen target VLM updates context and yields features for downstream action heads.

Claimed highlights:

• Flash Primitive Decoding for boxes, points, depth and 2D/3D trajectories;
• Lower error: ~28–61% lower 3D displacement error vs the AR target on ShareRobot-Bench / VABench-V;
• Speed: ~5× robot inference on Jetson Thor / DGX Spark, with even larger decoding speedups for driving;
• Downstream WAMs / diffusion action heads cross-attend VFlash plan features for higher control success.

Problem

Embodied models need not "think" only in text. Predicting where to grasp, how to move and how a scene evolves bridges perception and action — but plans must refresh often, and autoregressive coordinate tokens become the bottleneck.

Design

1. A tool call declares length; primitive type sets coordinate layout;
2. With frozen target-VLM context, a light (~5 non-causal Transformer layers) drafter trained on annotated geometry emits the full primitive in one parallel draft;
3. Boundary/position alignment matches training and decoding blocks to whole primitives and uses dedicated coordinate tokens or fixed-width fields;
4. Sub-task text still uses verified speculative decoding.

Results & caveats

The page reports gains on robot, dense-motion, detection and NAVSIM/Waymo driving setups; downstream control beats language-only or AR-plan baselines on several suites. Caveat: numbers are NVIDIA self-reported with custom TensorRT runtimes; this is a research/engineering system, not necessarily a consumer download.

Who it's for

• Researchers speeding up robot/AV visual planning and VLA inference;
• Engineers deploying Physical AI on Jetson Thor, DGX Spark and similar platforms;
• Product teams tracking System-1 vs System-2 splits in embodied AI.

Next steps

1. Read https://research.nvidia.com/labs/lpr/vflash/ ;
2. Watch LPR/GEAR for code and weight releases;
3. Cross-read with tracked NVIDIA Flora, NVIDIA ASPIRE and GR00T entries.