RAVEL: Asynchronous Rolling Inference for Flow-Based Vision-Language-Action Models

2026-09-28

AI Summary

This study addresses the high inference latency in flow-based vision-language-action (VLA) models caused by VLM encoding and multi-step iterative denoising. We propose RAVEL, an asynchronous inference framework that innovatively decouples VLM encoding from action generation. Specifically, it introduces a lightweight, fast observation pathway that directly conditions action experts to circumvent encoding bottlenecks. Furthermore, a rolling buffer mechanism is incorporated to execute imminent actions via single-step denoising while asynchronously precomputing future actions. Experiments demonstrate that this approach significantly reduces response latency in both simulated and real-world robotic tasks while preserving original performance, thereby enabling high-frequency, low-latency closed-loop control.

Abstract

Flow-based vision-language-action (VLA) models are highly effective for generalist robot manipulation, yet their reliance on computationally expensive VLM encoding and multi-step iterative action generation imposes a significant latency bottleneck. The resulting inference latency makes it difficult for robots to respond quickly, especially in dynamic environments. We address this limitation with RAVEL (Rolling Asynchronous VLA Enabling Low-Latency Control), an asynchronous inference framework that addresses the computational bottlenecks of both the VLM backbone and the action expert. To reduce the delay from multi-step action denoising, RAVEL allows near-term actions to be executed after a single denoising step by carrying partially denoised future actions forward in a rolling buffer. To avoid blocking on slow VLM encoding, RAVEL decouples VLM encoding from rolling action generation, allowing the action expert to operate continuously using the latest available VLM context, while a lightweight Fast Observation Pathway (FOP) directly conditions the action expert on current observations. Across simulated and real-world manipulation tasks, RAVEL consistently achieves substantially lower response latency while maintaining the task capability of the underlying VLA, enabling high-frequency and responsive closed-loop control.

Problem

• Vision-Language-Action models
• inference latency
• robot manipulation
• flow-based models
• real-time control

Innovation

• Asynchronous Inference
• Flow-Based VLA
• Rolling Buffer
• Decoupled Encoding
• Fast Observation Pathway