arXiv cs.AI· Weijia Liufu, Xiaoyu Guo, Ruiyi Chen, Jingzhi Liu, Kaidong Zhang, Xiwen Liang, Jianqi Lin, Dawei Sun, Yuze Wang, Rongtao Xu, Bingqian Lin, Bowen Yang, Tongtong Cao, Bowen Peng, Dongyu Zhang, Guangrun Wang, Min Wang, Liang Lin, Xiaodan Liang·· 13 小时前AI 评分44
RePO-VLA：面向视觉-语言-动作模型的恢复驱动策略优化框架
RePO-VLA: Recovery-Driven Policy Optimization for Vision-Language-Action Models
AI 导读
RePO-VLA 是一个恢复驱动策略优化框架，将成功、恢复与失败轨迹赋予不同角色，通过 Recovery-Aware Initialization、Progress-Aware Semantic Value Function 和 Value-Conditioned Refinement 学习从失败状态回归成功流形。
正文 · 原文
Authors:Weijia Liufu, Xiaoyu Guo, Ruiyi Chen, Jingzhi Liu, Kaidong Zhang, Xiwen Liang, Jianqi Lin, Dawei Sun, Yuze Wang, Rongtao Xu, Bingqian Lin, Bowen Yang, Tongtong Cao, Bowen Peng, Dongyu Zhang, Guangrun Wang, Min Wang, Liang Lin, Xiaodan Liang
Abstract:Vision-Language-Action (VLA) models remain brittle in long-horizon, contact-rich manipulation because success-only imitation provides little supervision for execution drift, while failed rollouts are often discarded. We introduce RePO-VLA, a recovery-driven policy optimization framework that assigns distinct roles to success, recovery, and failure trajectories. RePO-VLA first applies Recovery-Aware Initialization (RAI), slicing each corrective suffix into a new episode so corrective actions depend on the current adverse state rather than the preceding failure. It then learns a Progress-Aware Semantic Value Function (PAS-VF), aligning spatiotemporal trajectory features with instructions and successful references. The resulting labels salvage useful failure prefixes via reliability decay, while low-value labels mark drift and terminal breakdowns, teaching differences among nominal, failed, and corrective actions. The data engine turns adverse states into planner-generated or human-collected corrective rollouts, teaching recovery to the success manifold. Value-Conditioned Refinement (VCR) trains the policy to prefer high-progress actions. At deployment, a fixed high-value token biases actions toward the learned success manifold without online failure detectors or heuristic retries. We introduce the Failure-Recovery Benchmark (FRBench), with standardized error injection and recovery-focused evaluation. Across simulated and real-world bimanual tasks, RePO-VLA improves robustness, raising adversarial success from 20% to 75% on average and up to 80% in scaled real-world trials. Code and video results are available at this https URL.
| Comments: | Accepted at NeurIPS 2026. Weijia Liufu and Xiaoyu Guo contributed equally. Xiaoyu Guo is the project leader. Liang Lin and Xiaodan Liang are co-corresponding authors. Updated to the camera-ready version |
| Subjects: | Robotics (cs.RO); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2605.09410 [cs.RO] |
| (or arXiv:2605.09410v2 [cs.RO] for this version) | |
| https://doi.org/10.48550/arXiv.2605.09410 arXiv-issued DOI via DataCite |
Submission history
From: Xiaoyu Guo [view email]
[v1]
Sun, 10 May 2026 08:24:05 UTC (6,328 KB)
[v2]
Mon, 26 Oct 2026 18:36:06 UTC (6,342 KB)
来源：arXiv cs.AI · arxiv.org