AI

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

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 07/10/2026, 11:18:24 UTC.

Briefings

A research team developed a two-step flow denoising method to reduce real-time vision-language-action model latency from 61.557 ms to 21.956 ms for garment-folding tasks. They evaluated six execution methods on a physical task, finding Legato performed best among training-based approaches while Temporal Smoothing led among training-free methods. The study proposes a distributed framework balancing inference, action, and robot-control rates to address timing gaps between model inference and robot execution.

Key points

• Two-step flow denoising reduced model inference time from 61.557 ms to 21.956 ms for garment-folding tasks.
• Legato performed best among training-based methods, while Temporal Smoothing led among training-free methods.

Context

Vision-language-action models face a timing gap between low-rate inference and high-rate robot execution. This research addresses the latency bottleneck in real-time VLA systems.

Limitations

Evaluation was limited to a single long-horizon physical garment-folding task with six execution methods compared against a single baseline. Results are based on end-to-end latency measurements and may not generalize to all robot systems or task types.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 27/10/2026, 07:58:16 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeigevru4rrd2ps46mcg32uumqjmxujijdxs4izkderw4drejgohhma