NVIDIA and Partners Propose VFlash: Accelerating Visual Planning Decoding for Physical AI

NVIDIA, in collaboration with several universities, has proposed VFlash, a visual planning decoding framework for Physical AI. By leveraging geometric priors to generate full trajectories and visual tokens in parallel, it eliminates the latency bottleneck of autoregressive token-by-token generation. On Jetson Thor, the framework speeds up the decoding stage by 12x to 42x, boosts overall robotic pipeline speed by 5.1x, and reduces average 3D displacement error by 42.6% in autonomous driving and robotic arm tasks.