NVIDIA VFlash speeds up visual planning for robots and reduces trajectory errors
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
Robots will be able to update spatial plans more frequently without losing accuracy: parallel coordinate generation removes part of the delay between observation and action.
Flash Primitive Decoding exploits the known structure of points, boxes and trajectories: a compact drafter proposes the entire set of coordinates in one parallel pass instead of generating tokens sequentially. Across three robotics datasets and three target-model variants, mean 3D displacement error fell by 42.6% relative to autoregressive decoding. Across four MolmoSpaces suites, an action model using VFlash plans completed an average of 70.2% of tasks, compared with 62.6% for the baseline configuration. On Jetson Thor and DGX Spark, the researchers measured 12-fold faster decoding for robot planning; driving tasks saw speedups of 11–49 times depending on the platform. The results were published as a preprint, and speed measurements used NVIDIA's specialized TensorRT pipelines.
Sources 2