NVIDIA introduced VFlash, a decoding framework for Physical AI. It enables parallel generation of visual primitives like boxes, trajectories and dense motion. This cuts 3D waypoint error by up to 61% versus autoregressive targets. It also reduces mean 3D displacement error by 42.6% across robot datasets. Robot inference runs five times faster on Thor and Spark platforms. Decoding latency drops 12 to 42 times on Jetson Thor, DGX Spark and H100. The method supports variable output lengths for tasks like NAVSIM and Waymo. It improves robot success rates from 62.6% to 70.2% in MolmoSpaces. Driving inference latency falls over three times on Jetson Thor and DGX Spark. The drafter predicts complete primitives in one pass without token-level agreement. Boundary and position alignment are critical for parallel prediction accuracy. The framework integrates with frozen VLMs and downstream action models. Code and technical report are publicly available.
What changed
Autoregressive generation of coordinate tokens made detailed visual plans expensive and slow.
What it unlocks
Generating complete visual primitives in parallel for faster and more accurate robot and autonomous driving inference.
- Up to 61% lower 3D waypoint error
- 5× faster robot inference on Thor and Spark
- 42.6% lower mean 3D displacement error
- 12–42× faster decoding latency
What you need to act on it
- a frozen target VLM
- a lightweight drafter trained on annotated geometry
- TensorRT-Edge-LLM runtimes and TensorRT engines
Sources