ProbeFlow: Training-Free Adaptive Flow Matching for Vision-Language-Action Models

TL;DR

ProbeFlow is a training-free adaptive inference framework that significantly reduces latency in vision-language-action models for robotic control by dynamically adjusting evaluation steps based on trajectory complexity.

Contribution

It introduces a novel, training-free method to adaptively schedule inference steps, greatly accelerating action decoding without sacrificing success rates.

Findings

Accelerates action decoding by 14.8x on MetaWorld

Reduces system latency by 2.8x without performance loss

Effectively handles long-horizon tasks and real-world deployments

Abstract

Recent Vision-Language-Action (VLA) models equipped with Flow Matching (FM) action heads achieve state-of-the-art performance in complex robot manipulation. However, the multi-step iterative ODE solving required by FM introduces inference latency that precludes responsive physical control. While current acceleration efforts optimize the Vision-Language Model (VLM) backbone, the action head bottleneck remains overlooked. To address this, we propose ProbeFlow, a training-free adaptive inference framework tai- lored for continuous robotic control. By evaluating geometric trajectory complexity via the cosine similarity between initial and lookahead velocity vectors, ProbeFlow dynamically sched- ules integration steps to prune redundant network evaluations. On the MetaWorld benchmark, it accelerates action decoding by 14.8x (reducing average steps from N = 50 to 2.6) and cuts end-to-end system latency by 2.8x without compromising the manipulation success rate. On the long-horizon LIBERO benchmark, the probe automatically allocates a denser schedule to navigate semantic bottlenecks, effectively resolving the flow solver delay. Real-world physical deployments confirm that ProbeFlow successfully mitigates action decoding latency while ensuring execution stability, offering a highly practical solution for low-latency continuous generative policies.