外观
论文arXiv 2610.2138913 天前
Fast-VLA：机器人操作中只需思考一次
原文链接：Fast-VLA: You Only Think Once for Robot Manipulation（arXiv 2610.21389 · 发布日期：2026-10-15）
Fast-VLA: You Only Think Once for Robot Manipulation
Fast-VLA通过每回合一次调用冻结4B VLM构建任务记忆，在RoboTwin和LIBERO评测中验证机器人操作策略加速。
核心要点
- 动机：现有VLA在反馈环中反复重算任务上下文，代价高且上下文变化慢。
- 方法：每回合一次调用冻结4B VLM，将指令和初始视图编码为固定多层键值任务记忆。
- 执行：轻量在线actor融合当前视觉与本体反馈，检索任务记忆并预测动作块。
- 结果：在四项评测获79.11%至95.95%成功率，推理时间降10.9-17.1x；提供项目页。
验证范围与局限：验证范围为RoboTwin 2.0 Clean/Randomized、LIBERO与LIBERO-Plus评测；摘要未说明真机实验。
英文摘要 (Abstract)
Vision-language-action (VLA) policies commonly place a multi-billion-parameter model inside the feedback loop, repeatedly recomputing task context whenever they predict a new action chunk. Yet task understanding changes far more slowly than the robot state. We introduce Fast-VLA, which invokes a frozen 4B VLM once per episode to contextualize the instruction and initial views as fixed, multi-layer key-value task memory. During execution, a lightweight online actor fuses learned time- and arm-indexed queries with current visual and proprioceptive feedback to form action queries. These queries retrieve task context from the fixed memory and predict action chunks without re-running the foundation model. Fast-VLA achieves 92.46% and 92.40% success on RoboTwin 2.0 Clean and Randomized, and 95.95% and 79.11% on LIBERO and LIBERO-Plus evaluation. Against three continually conditioned baselines on the same hardware, it reduces RoboTwin amortized policy inference time by 10.9-17.1x and requires 1.22 GiB of post-prefill online resident memory. Task-memory swaps further show that the actor relies on episode-matched task memory. These results show that task context can be established once while closed-loop control remains responsive throughout execution. Project page: https://garfield-kh.github.io/FastVLA
作者：Kehong Gong, Guanli Hou, Xiaoxu Zheng, Qi Wang, Mingxi Xu, Anbang Zhai, Xiaoyu He, Mingyuan Zhang, Michael Bi Mi
来源
本页事实对应的公开页面，以原文为准。