Guided Action Flow - Q-Guided Inference for Flow-Matching Vision-Language-Action Policies

Sep 21, 2026

Uploaded by Eliana

Flow-matching vision-language-action policies generate robot action chunks via an iterative transport process, enabling test-time guidance without retraining the underlying policy. Guided Action Flow presents an inference-time framework that keeps a pretrained SmolVLA policy frozen and learns an action-chunk critic to guide its reverse-time flow sampler. The critic is trained from real success and failure rollouts, conditions on task-description features from the frozen language pathway, and is used only through action gradients during sampling. Evaluations on LIBERO show substantial success gains from single-task and multi-family critics, while highlighting critic generalization and uncertainty-aware guidance as key bottlenecks.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages9

Reading Time23 min

AccessFree to download

Introduction

Problem and motivation

Modular adaptation with a critic

Connection to guidance and Q-guided flow

Guided Action Flow: Q-Guided Inference for Flow-Matching Vision-Language-Action Policies Liuhaichen Yang 1 , Zhuang Jiang 1 , Chenchao Sheng2 , Zezhi Tang 1∗ arXiv :2607 .02092v2 [ cs .RO] 3 Jul 2026 Abstract—Flow-matching vision-language-action policies generate robot action chunks through an iterative transport process, creating an opportunity for test-time guidance without retraining the base policy. We study this opportunity in Guided Action Flow, an inference-time framework that keeps a pretrained SmolVLA policy frozen and uses a learned action-chunk critic to guide its reverse-time flow sampler. The critic is trained from real success and failure rollouts, can condition on task-description features from the frozen SmolVLA language pathway, and is used only through action gradients during sampling. We evaluate the approach on LIBERO manipulation tasks. A single-task critic improves success from 68.0% to 82.0% on one seed window and from 82.0% to 86.0% on another. A multi-family task-description critic improves validation success from 46.0% to 56.0%, while the locked held-out test gain is positive but modest, from 65.0% to 67.5% . These results support the feasibility of Q-guided inference for frozen flow-matching VLA policies, while showing that critic generalization and uncertainty-aware guidance remain the central bottlenecks. Index Terms—vision-language-action models, robot manipulation, flow matching, test-time guidance, offline reinforcement learning I. INTRODUCTION Vision-language-action (VLA) policies have become a practical route to language-conditioned robot manipulation. Earlier language-conditioned manipulation systems already showed that semantic task descriptions can improve generalization beyond fixed goal labels, using architectures that combine pretrained vision-language features, imitation learning, or 3D action representations [1]–[3] . Recent VLA systems extend this direction by scaling policy learning with larger robot datasets, stronger language-vision backbones, and sequence-generation architectures. Systems such as RT-1, RT-2, PaLM-E, Open XEmbodiment, Octo, OpenVLA, π0 , FAST, and SmolVLA scale policy learning by combining language conditioning, large robot datasets, and sequence-generation architectures [4]–[12] . This progress is especially relevant for benchmarks such as LIBERO, LIBERO-Plus, and LIBERO-PRO, which expose the difficulty of robust language-conditioned manipulation under task, object, layout, and evaluation distribution shifts [13]–[15] . Recent embodied interfaces, including SignVLA, further broaden the language and multimodal instruction channels through which robots may receive tasks [16] . The resulting policies are increasingly capable, but their adaptation story is still incomplete. A pretrained VLA may 1 Department of Computer Science, University College London. 2 Department of Mechanical Engineering, University College London. Correspondence: [zezhi.tang@ucl.ac.uk](zezhi.tang@ucl.ac.uk) execute many benchmark tasks, yet still fail when the task distribution shifts, when early actions create compounding errors, or when multiple plausible action chunks satisfy the language instruction locally but only one leads to task completion. The obvious response is to fine-tune the VLA. In practice, however, full-policy fine-tuning can be expensive, hardwaresensitive, and difficult to validate when the user only has a small amount of task-local data. This is especially true for consumer-GPU settings, which motivated the use of compact SmolVLA checkpoints in this project. This paper studies a more modular adaptation path. Instead of updating the base policy, we train a smaller critic from rollouts of the frozen policy and use that critic only at inference time. The critic does not choose among a finite set of proposed actions, and it is not used to fine-tune the VLA. It provides a gradient over the continuous action chunk that biases the policy sampler toward highe

FAQs

What is Guided Action Flow and what problem does it address?

Guided Action Flow is an inference-time adaptation method that keeps a pretrained SmolVLA policy frozen while using a learned action-chunk critic to guide the reverse-time flow sampler, aiming to improve success under task distribution shifts.

How is the critic trained and how is it used during inference?

The critic is trained from real success and failure rollouts, can condition on task-description features from the frozen SmolVLA language pathway, and is used only through action gradients during sampling.

How do the results on LIBERO validate the approach?

A single-task critic improves success from 68.0% to 82.0% on one seed window and from 82.0% to 86.0% on another; a multi-family critic improves validation success from 46.0% to 56.0%, with a modest positive gain on held-out test results from 65.0% to 67.5%.

Sep 21, 2026

Uploaded by Eliana

Guided Action Flow - Q-Guided Inference for Flow-Matching Vision-Language-Action Policies_Page 1

Open in WPS to view all 9 pages