Science & Tech · AI · 2 hrs ago
NVIDIA’s VFlash speeds robot trajectory predictions and reduces errors
NVIDIA researchers developed VFlash, a method for helping robots predict movements and other spatial information.
Conventional vision-language models often produce coordinates one at a time, which can slow predictions as trajectories become longer or more detailed.
VFlash instead generates a complete trajectory in one pass, using a lightweight model to predict its coordinates together.
In robot-manipulation tests, it cut prediction time by about 80% on NVIDIA’s Jetson Thor and DGX Spark platforms.
The researchers also found that VFlash reduced 3D trajectory errors by up to 61% in some tests.
When its predictions were used for robot-control tasks, the system’s average success rate was 70.2%, compared with 62.6% for the baseline.
The researchers also tested the method for autonomous driving, but no further deployment plans are stated.
NVIDIA researchers developed VFlash, a method that generates complete robot trajectories and other spatial predictions in one pass.
In specific tests, it decoded robot trajectories up to 12 times faster and reduced 3D trajectory prediction errors by as much as 61%.
Robot-manipulation inference latency fell by about 80% on NVIDIA’s Jetson Thor and DGX Spark platforms.
On the MolmoSpaces benchmark, a system using VFlash trajectories achieved a 70.2% average success rate, compared with 62.6% for the baseline.
The research also tested VFlash for 3D point-motion forecasting and autonomous-driving trajectory generation.
- Who
- NVIDIA researchers and collaborators developed VFlash.
- What
- The method generates visual predictions, including robot trajectories, in a single pass. Tests found faster decoding, lower prediction errors and improved manipulation success rates.
- When
- The research was published in November 2026.
- Where
- Tests included NVIDIA’s Jetson Thor and DGX Spark platforms, as well as the MolmoSpaces benchmark.
- Why
- Conventional vision-language models often generate spatial predictions one coordinate at a time, which can increase processing time as predictions grow more complex.
This story does not have two clearly opposing sides.
This story does not have a timeline yet.
- Trajectory decoding speed
- Up to 12 times faster
- 3D prediction error
- Reduced by as much as 61% in specific tests
- Jetson Thor inference latency
- 3,552.7 milliseconds to 703.3 milliseconds
- DGX Spark inference latency
- 3,356.5 milliseconds to 649.9 milliseconds
- MolmoSpaces average success rate
- 70.2% with VFlash trajectories, compared with 62.6% for the baseline
- Publication
- November 2026