arXiv Robotics — research abstracts· Wenxuan Guo, Yingping Liang, Zhaoguo Wang, Ziyuan Li, Angyuan Ma, Zhenyu Wu, Chengkun Wang, Bingyao Yu, Ze Chen, Borui Zhang, Xiuwei Xu, Meng Zhang, Erjin Zhou, Jiwen Lu·2026-10-14 04:00 UTC· 2 days agoSignalEditorial score85

LoopWAM: Closed-Loop Generalist Policy Improvement from Deployment Rollouts

Summary

This paper presents LoopWAM, a closed-loop world-action model that learns directly from full-quality deployment rollouts, including successes, suboptimal behavior, and failures. By integrating source-dependent noise sampling, quality steering, and quality-weighted supervision, LoopWAM achieves significant improvements in success rates on benchmarks like RoboChallenge Table30-V2, RoboTwin 2.0, and RoboCasa365, demonstrating effective closed-loop policy improvement under both multi-task learning and single-task post-training scenarios.

Editorial context

LoopWAM introduces a closed-loop world-action model that improves robot policies by learning from mixed-quality deployment rollouts, achieving significant success improvements on multiple benchmarks without additional embodied pretraining.

Source: arXiv Robotics — research abstracts · Read original article ↗

Loading article text…

Source：arXiv Robotics — research abstracts · arxiv.org

#Evaluation #Robot policies #Imitation learning #Reinforcement learning