Two-step flow denoising cuts VLA inference from 61.6ms to 22ms for real-time robot control
Di Wu · hf · 2026-10-07
The paper characterizes the timing gap between low-rate VLA inference and high-rate robot execution: repeated Flow Matching denoising dominates inference cost, while robot-side delays stem from perception acquisition, communication scheduling, and physical response.
Observing that the velocity field is stable in early integration with stronger directional correction near terminal steps, the authors propose two-stage non-uniform denoising, cutting steps from 10 to 2 and model inference from 61.557 ms to 21.956 ms. They also build a distributed real-time VLA framework with independent inference, action-publication, and control rates.
Using π0.5 as baseline on a long-horizon garment-folding task, they evaluate six real-time execution methods: Legato leads training-based methods, Temporal Smoothing leads training-free ones; combining two-step denoising with these methods substantially reduces inference cost with only minor performance loss.
More from Embodied
- 2023's ACT + ALOHA was the original physical AI era, says dev — lukas_m_ziegler · 2026-10-07
- EAPN: execution-aligned noise fixes mode switching in asynchronous replanning, 96.7% bimanual success — Di Wu · 2026-10-07
- CTP: single-pass multimodal robot policy hits 97.25% on LIBERO, cuts inference latency to 75.8ms — Di Wu · 2026-10-07
- Magic-W0: a structured world-action foundation model topping RoboDojo-Sim at 27.10 — Xuhua Chen · 2026-10-07
- RaiSim is back: v2.8.0 brings ultra-fast accurate CPU simulation plus photorealistic rendering — ChongZzZhang · 2026-10-07
- LiteReality-Agent turns images into simulation-ready 3D scenes for MuJoCo robot planning — elliottszwu · 2026-10-07