Toward Real-Time VLAs: Stage-Aware Two-Step Flow Denoising and System-Level Evaluation
Two-step flow denoising cuts VLA inference from 61.6 ms to 22.0 ms with a small performance drop.
The paper measures the timing gap between low-rate vision-language-action inference and high-rate robot execution, finding repeated flow-matching denoising is a major inference cost. A stage-aware two-step denoiser reduces steps from 10 to 2 and inference time from 61.557 ms to 21.956 ms. Using pi0.5, six real-time execution methods are tested on long-horizon garment folding; Legato is strongest among training-based methods and Temporal Smoothing among training-free ones. Combining two-step denoising with those methods cuts inference cost with only a small drop in task performance.
- Non-uniform two-stage denoising cuts flow-matching steps from 10 to 2.
- Model inference falls from 61.557 ms to 21.956 ms.
- Six real-time methods are compared on long-horizon physical garment folding using pi0.5.
- Legato leads training-based methods; Temporal Smoothing leads training-free methods.
Full article197 words · extracted from huggingface.co · click to collapse
Vision-language-action (VLA) models face a timing gap between low-rate inference and high-rate robot execution. We characterize this gap through end-to-end latency measurements of model inference and the robot execution chain. Repeated Flow Matching denoising contributes substantially to inference cost, while robot-side delays mainly arise from perception acquisition, communication scheduling, and physical response. Analysis of the velocity field shows relatively stable magnitude and direction in early integration, followed by stronger directional correction near the terminal steps. Based on this stage heterogeneity, we propose two-stage non-uniform denoising, reducing the number of steps from 10 to 2 and model-inference time from 61.557 ms to 21.956 ms. We also develop a distributed real-time VLA framework with independent inference, action-publication, and robot-control rates, modular observation acquisition, and action-provenance logging. Using π0.5 as the baseline, we evaluate six real-time execution methods on a long-horizon physical garment-folding task. Legato performs best overall among training-based methods, while Temporal Smoothing leads among training-free methods; both perform strongly in task success, completion time, action continuity, and acceleration smoothness. Combining two-step denoising with representative execution methods substantially reduces inference cost with a small reduction in task performance. These results motivate joint optimization of model-inference efficiency and robot-system timing.
Text extracted automatically; images, tables and formatting may be missing. Original: https://huggingface.co/papers/2609.39822