arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2610.24573cs.RO
收藏

针对冻结VLA策略的检索触发动作投影调整:实证研究与失败分析

Retrieval-triggered Action Projection Adjustment for Frozen VLA Policies: An Empirical Study and Failure Analysis

Hainan Bielefeld University of Applied Sciences(海南比勒费尔德应用科学大学)

Jinming Wang, Ruiheng Zhang, Yong Wang, Zipan Xiong, Keyuan Luo

AI总结:

本研究针对冻结VLA策略提出检索触发投影校正(RTPC),经CALVIN任务评估发现其可降低近碰撞事件率但未在独立复制中重现,分析指出相关局限性,为该类干预提供实证评估与失败分析。

摘要:

冻结的视觉-语言-动作(VLA)策略无法在不更新底层模型的情况下纳入此前遇到的失败案例。我们提出检索触发投影校正(RTPC),这是一种用于冻结OpenVLA策略的可审计外部内存包装器,它检索应力收集的局部深度记录,并应用仅平移的投影校正,同时保持旋转和夹爪命令不变。在CALVIN单任务操纵任务上,我们使用主要队列和独立复制队列对RTPC进行评估。在主要评估中,RTPC降低了基于深度的近碰撞或停滞代理事件率,且未改变任务成功次数。然而,这一降低效果未在独立复制中重现,且检索到的校正方向与随机同家族方向无差异。分析指出端点耦合、批次敏感性以及缺乏与触发匹配的无内存基线是将效果归因于基于检索的校正的限制因素。这些发现未确立物理安全性提升或特定内存优势,但为冻结VLA策略的检索条件干预提供了实证评估与失败分析。

英文摘要:

Frozen Vision-Language-Action (VLA) policies cannot incorporate previously encountered failures without updating the underlying model. We present retrieval-triggered projection correction (RTPC), an auditable external-memory wrapper for frozen OpenVLA policies that retrieves stress-collected local-depth records and applies translation-only projection correction while keeping rotation and gripper commands unchanged. RTPC is evaluated on CALVIN single-task manipulation using primary and independent replication cohorts. In the primary evaluation, RTPC reduced a depth-based near-collision or stall proxy event rate without changing task-success count. However, this reduction was not reproduced in the independent replication, and retrieved correction directions were not distinguishable from random same-family directions. Analysis identifies endpoint coupling, batch sensitivity, and the lack of a trigger-matched memory-free baseline as limitations for attributing effects to retrieval-based correction. These findings do not establish physical safety improvement or memory-specific benefits, but provide an empirical evaluation and failure analysis of retrieval-conditioned intervention for frozen VLA policies.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑