Toward Real-Time VLAs: Stage-Aware Two-Step Flow Denoising and System-Level Evaluation

Toward Real-Time VLAs. Stage-Aware Two-Step Flow Denoising and System-Level Evaluation. By Di Wu and ten others, from Magiclab Robotics, Zhejiang University, and four other institutions. Published on the third of October. Its code repository is worked out: about one hundred and forty source files, with a readme and a licence.

Vision-language-action models run inference at a low rate, while robots execute at a high rate, and the two do not line up. The authors characterise this timing gap by measuring end-to-end latency across both model inference and the robot's execution chain. On the model side, repeated Flow Matching denoising contributes substantially to the cost. On the robot side, delays come mainly from three places: acquiring perception, scheduling communication, and the physical response of the machine.

The model-side fix starts from the velocity field. Its analysis shows that in early integration the magnitude and direction stay relatively stable, and that stronger directional correction arrives only near the terminal steps. Because the stages behave differently, they need not be treated alike. The proposal is two-stage non-uniform denoising, which cuts the number of denoising steps from ten to two.

That cut brings model-inference time down from sixty-one point five five seven milliseconds to twenty-one point nine five six milliseconds.

For the robot side, the authors develop a distributed real-time VLA framework. Inference, action publication and robot control each run at their own independent rate. Observation acquisition is modular, and the framework logs the provenance of every action, so the system's timing can be inspected rather than guessed at.

With pi zero point five as the baseline, six real-time execution methods are evaluated on a long-horizon physical task: folding garments. Among methods that need training, Legato performs best overall. Among training-free methods, Temporal Smoothing leads. Both perform strongly on task success, completion time, action continuity, and acceleration smoothness. Combining two-step denoising with representative execution methods substantially reduces inference cost, at the price of a small reduction in task performance.

Neither half of the problem can be solved alone. A faster model still waits on perception, communication scheduling and physical response, and a well-timed robot still waits on the model. The results argue for optimising model-inference efficiency and robot-system timing jointly, as one problem.

The paper's contributions line up with that view. On the model side, a reading of the velocity field turns into a non-uniform denoiser built around the two stages the field exhibits. On the system side, a framework separates the rates at which the robot thinks, publishes and moves. And the garment-folding comparison puts several execution methods side by side on one long task, so the trade between speed and success can be seen directly.