用任务提示引导冻结模型，让机器人更准完成动作。
Guided Action Flow: Q-Guided Inference for Flow-Matching Vision-Language-Action Policies
- 推理时用任务感知梯度引导动作流采样
- 真实机器人任务成功率从47.5%提升至85%
- 无需微调模型，适合快速部署到新场景
在特定机器人和工作空间上部署预训练的流匹配视觉-语言-动作（VLA）策略通常需要任务定制化调整，而全模型微调成本高且会改变基础行为。本文提出引导动作流（QGF），在推理阶段保持预训练的SmolVLA策略冻结，通过任务特异性的动作片段评判器提供的梯度来引导其逆时序动作流采样。QGF使用100次真实机器人回放数据，采用离线隐式Q学习训练一个视觉Transformer评判器与价值模型。该评判器基于机器人状态、冻结的双摄像头SmolVLA视觉标记以及策略的归一化50步动作片段进行条件判断。在真实机器人水瓶放置任务中，当η=2时，成功率从40次中的19次（47.5%）提升至34次（85.0%），超时次数从13次降至3次。加入黄色卷尺作为视觉干扰物后，QGF成功完成12次中的6次，而冻结的SmolVLA仅完成11次中的0次。结果表明，部署回放数据可为冻结的流匹配VLA提供有效的动作空间引导信号，无需策略微调即可显著提升性能。
原文摘要 · Abstract (English)
Deploying a pretrained flow-matching vision-language-action (VLA) policy on a particular robot and workspace often calls for task-specific adaptation, while full- policy fine-tuning is costly and changes the base behavior. We present Guided Action Flow, an inference-time method that keeps a pretrained SmolVLA policy frozen and steers its reverse-time action-flow sampling with gradients from a task-specific action-chunk critic. QGF trains a visual Transformer critic and value model with offline Implicit Q-Learning on 100 real-robot rollouts. The critic conditions on robot state, frozen dual-camera SmolVLA visual tokens, and the policy's normalized 50-step action chunk. On a real-robot water-bottle placement task, QGF with \b{eta} = 2 increases success from 19/40 episodes (47.5%) to 34/40 episodes (85.0%) and reduces timeouts from 13 to 3. With a yellow tape measure added as a visual distractor, QGF completes 6/12 episodes, compared with 0/11 for frozen SmolVLA. These results show that deployment rollouts can provide an effective action-space guidance signal for improving a frozen flow-matching VLA without policy fine-tuning.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。