NVIDIA Research has published VFlash, a method that lets a robot’s vision-language model write out a complete movement plan in one step instead of building it a piece at a time. On a Jetson Thor, the company’s robotics computer, the project page reports roughly five times faster robot inference. Plans also came out more accurate than the slow method produced.
Language models normally write one small piece of output at a time, and each piece waits for the one before it. That is tolerable for a sentence. A robot path is a long list of coordinates, and every coordinate takes several pieces. NVIDIA’s own examples show how fast this grows: Waymo’s driving benchmark asks for 20 waypoints over 5 seconds, and the Alpamayo driving system needs 64 waypoints spread across 6.4 seconds.
VFlash borrows an idea from text models called speculative decoding, where a cheap model drafts and an expensive one checks. Here the drafter is small, with five transformer layers, and it guesses a whole path in one parallel pass. The large vision model stays frozen, then reads the finished plan to update what it knows. The drafter can do this because a path has a known shape. Every position in the output already has a job (which waypoint, which axis) before anything gets filled in.
The speed claim comes with specifics. According to NVIDIA Research, one-at-a-time writing eats 75 to 87 percent of the model’s processing time on Jetson Thor, and VFlash shrinks that stage by 12 to 42 times. In one robot pipeline, decoding dropped from 3,107.6 milliseconds to 258.2, and the whole model pipeline went from 3,552.7 milliseconds to 703.3. The measurements cover one request at a time and exclude the sensors and motors, so they describe the model’s thinking time, not the robot’s full reaction time.
Accuracy is the more surprising result. Tested on three robot datasets with three versions of the target model, the average 3D position error fell by 42.6 percent compared with the same model writing its plan the slow way. The researchers attribute this to training the drafter directly on annotated geometry, so it never has to imitate the big model’s exact output.
Better plans did translate into better robot behavior in simulation. In the MolmoSpaces benchmark, a control model built on NVIDIA’s Cosmos-Predict2.5 succeeded 70.2 percent of the time using VFlash plans, against 62.6 percent for the baseline. Plans from the slow method scored 61.0 percent, slightly below having no plan at all. A plan that arrives late or wobbles can hurt more than it helps.
Driving results are less dramatic. On the NAVSIM benchmark, VFlash scored 90.7, the same as Qwen-Drive-1.0-RL in NVIDIA’s comparison table. On Waymo, its 3-second position error was 1.22 meters against 1.19 for Qwen-Drive, and its rating score was 7.86 against 7.91. NVIDIA describes this as competitive performance with much lower decoding time, and the table supports that wording but nothing stronger.
Everything above was measured by the authors, who come from NVIDIA, Johns Hopkins University, Hong Kong Polytechnic University, Purdue University, and the University of Hong Kong. The speed figures use NVIDIA’s customized runtimes on NVIDIA hardware, and the robot success rates come from simulation. The page lists no tests on a physical robot.
A planning pass that falls from about 3.6 seconds to 0.7 seconds on a Jetson Thor is the gap between a robot that replans a few times per task and one that can recheck its plan between movements.
Reported by NVIDIA Research on its VFlash project page, “Enabling System-1 VLMs for Physical AI” (no publication date given).
