NVIDIA Unveils VFlash for 12x Faster Robot Trajectory Decoding
Overview
NVIDIA researchers published VFlash in November 2026, claiming up to 12x faster decoding and 61% lower error on Jetson Thor/DGX Spark. A new arXiv abstract later cites ~25x speedup for 3D forecasting and higher task success rates.
Generated from attributed reports · Updated 2 days ago
Event evidence and corrections
1 attributed source owners. Ownership does not establish independent confirmation. Quantities are reported separately and are never added together.
Reported quantity · inference speedup: 12 other · Basis not reported
“VFlash enables up to 12x faster action decoding and a 5.1x overall robot inference speedup on edge GPUs like the Jetson Thor and DGX Spark. What’s surprising is that it ends up even more accurate than the VLM backbone, with up to 61% lower 3D trajectory error. The intuition is that VFlash avoids the cascading errors of autoregressive decoding, where every coordinate is written in ink and every later point loses the c”
Exact source · revision 1nvidia
Reported quantity · overall robot inference speedup: 5.1 other · Basis not reported
“VFlash enables up to 12x faster action decoding and a 5.1x overall robot inference speedup on edge GPUs like the Jetson Thor and DGX Spark. What’s surprising is that it ends up even more accurate than the VLM backbone, with up to 61% lower 3D trajectory error. The intuition is that VFlash avoids the cascading errors of autoregressive decoding, where every coordinate is written in ink and every later point loses the c”
Exact source · revision 1nvidia
Reported quantity · bounding box token count: 6 other · Basis not reported
“Most embodied prediction tasks like object localization, robot trajectory planning, and dense 3D motion are fixed-length float vectors. For example, a bounding box is always 6 tokens, a 10-waypoint robot arm trajectory is 54: we know these beforehand. Instead of decoding one token at a time, we can output the entire geometric shape in one go.”
Exact source · revision 1nvidia
Reported quantity · robot arm trajectory token count: 54 other · Basis not reported
“Most embodied prediction tasks like object localization, robot trajectory planning, and dense 3D motion are fixed-length float vectors. For example, a bounding box is always 6 tokens, a 10-waypoint robot arm trajectory is 54: we know these beforehand. Instead of decoding one token at a time, we can output the entire geometric shape in one go.”
Exact source · revision 1nvidia
Reported quantity · inference latency on Jetson Thor: 703.3 other · Basis not reported
“al information, this sequential approach can increase processing time. VFlash instead uses the known structure of visual predictions to generate an entire geometric object or trajectory in one forward pass. In tests involving robot manipulation, the researchers reported that VFlash reduced inference latency from 3,552.7 milliseconds to 703.3 milliseconds on NVIDIA’s Jetson Thor platform. This represents an approximat”
Exact source · revision 2Source owner not reported
Reported quantity · inference latency on DGX Spark: 649.9 other · Basis not reported
“ely 80% reduction compared with conventional autoregressive decoding. On DGX Spark hardware, latency fell from 3,356.5 milliseconds to 649.9 milliseconds. Model predicts complete trajectories VFlash works by identifying the structure of a requested visual prediction before generating its coordinates. For example, a bounding box has a fixed coordinate layout, while a trajectory can contain a variable number of two-dim”
Exact source · revision 2Source owner not reported
Reported quantity · Reported success rate: 70.2 percent · Basis not reported
“or robotic manipulation. The model achieved an average success rate of 70.2% across four task suites, compared with 62.6% for the baseline system. The largest improvements appeared in pick-and-place and object-closing tasks, where success increased by 11.6 and 12.8 percentage points, respectively. Beyond robotics, the researchers tested VFlash for autonomous driving using NAVSIM and Waymo datasets. The model generate”
Exact source · revision 2Source owner not reported
Reported quantity · generation speedup: 25 other · Basis not reported
“pproximately 25x faster generation for 3D point motion forecasting. 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.”
Exact source · revision 1Source owner not reported
Reported quantity · forward pass count: 1 other · Basis not reported
“We introduce VFlash, a method for fast, accurate visual planning through Flash Primitive Decoding. VFlash exploits the known structure of visual primitives to predict their coordinates jointly, using a lightweight drafter that generates an entire primitive in one forward pass. The layout adapts to the primitive's type, dimensionality, and length, supporting compact boxes and long, variable-length trajectories. Cruci”
Exact source · revision 1Source owner not reported
Developments
- 2026-11-06 05:00 UTC · 1 reportsVFlash: Fast, Accurate Visual Planning via Flash Primitive DecodingarXiv Robotics — research abstracts：VFlash: Fast, Accurate Visual Planning via Flash Primitive Decoding
- 2026-11-05 16:13 UTC · 2 reportsVFlash: Parallel Visual Parsing for Physical AIJim Fan — X Robotics：The era of System One Vision Model is here. Introducing VFlash: parsing the visual world in parallel at unmatched speeds for Physical AI
Report timeline
Follow attributed reports and material updates.
- arXiv Robotics — research abstractsSignalVFlash: Fast, Accurate Visual Planning via Flash Primitive Decoding
This research introduces VFlash, a method for fast and accurate visual planning in physical AI tasks like robotics and autonomous driving. By exploiting the structure of visual primitives, VFlash predicts coordinates jointly using a lightweight drafter, supporting compact boxes and variable-length trajectories without requiring exact token sequence reproduction. Evaluated on 2D and 3D robot trajectory tasks, VFlash reduces prediction errors relative to autoregressive decoding. Preliminary measurements indicate approximately 25x faster generation for 3D point motion forecasting in single-rollout scenarios. 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.
- Jim Fan — X RoboticsSignalThe era of System One Vision Model is here. Introducing VFlash: parsing the visual world in parallel at unmatched speeds for Physical AI
Jim Fan announces VFlash, a model that enables parallel decoding of geometric shapes for embodied AI tasks. It claims up to 12x faster action decoding and 5.1x overall inference speedup on edge GPUs like Jetson Thor and DGX Spark, with up to 61% lower 3D trajectory error compared to autoregressive methods. The post states the approach is open paper and open source.
- Robotics — Robot learning news discoverySignalNVIDIA's VFlash speeds robot trajectory decoding 12x, cuts prediction errors by 61%
NVIDIA researchers have developed a new visual planning method called VFlash that accelerates robot trajectory decoding by up to 12 times while reducing 3D trajectory prediction errors by as much as 61% in specific tests. The method allows vision-language models to generate complete trajectories in a single pass, cutting inference latency by about 80% on NVIDIA's Jetson Thor and DGX Spark platforms. In robot manipulation tests, VFlash reduced average displacement error significantly compared with conventional autoregressive approaches. The research was published in November 2026.
Timezone · UTC
Article dates follow your selected timezone. Briefing editions use Hong Kong time (UTC+8).