LoopWAM: Closed-Loop Generalist Policy Improvement from Deployment Rollouts (arXiv 2610.18873)

Forum topic · 小凯 · 2026-10-15

Summary

LoopWAM is a closed-loop world-action model that enables generalist robot policies to improve from deployment rollouts rather than relying solely on expert demonstrations. Real-world trajectories contain successes, suboptimal behaviors, and failures; naively imitating all actions can reinforce poor behavior, while filtering out failures discards evidence of physical interaction consequences. LoopWAM learns from this full quality spectrum using rollouts from itself and other deployed policies. Source-dependent noise sampling unifies behavior learning and action-conditioned world modeling, while quality steering and quality-weighted supervision retain full video supervision and regulate action imitation. Video and action tokens interact throughout parallel denoising, letting learned consequence information directly guide action generation. On RoboChallenge Table30-V2, three deploy-train cycles raise success rate from 23.67% to 43.33% (+19.7 points), with a final score of 56.80. LoopWAM also achieves 94.5% average success on RoboTwin 2.0 and 49.5% overall success on RoboCasa365 without extra embodied pretraining, ranking best overall across three benchmarks. On a challenging shirt-folding task, two rounds of deployment-driven improvement lift success from 13.20% to 55.56% after single-task post-training.

Paper Overview

Field: Computer Vision / Robotics Authors: 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 Published: 2026-10-13 arXiv: 2610.18873

Abstract

Scaling expert demonstrations alone does not enable generalist robot policies to improve from deployment. Real-world rollouts contain successes, suboptimal behavior, and failures: imitating every recorded action can reinforce poor behavior, while filtering out failures discards evidence of physical interactions and their consequences. We introduce LoopWAM, a closed-loop world-action model that learns directly from this full quality spectrum, using rollouts from both itself and other deployed policies.

Source-dependent noise sampling unifies behavior learning and action-conditioned world modeling, while quality steering and quality-weighted supervision retain full video supervision across qualities and regulate action imitation. Video and action tokens interact throughout parallel denoising, allowing learned consequence information to directly guide action generation.

Results

RoboChallenge Table30-V2: three deploy-train cycles improve success rate from 23.67% to 43.33% (+19.7 percentage points), with a final score of 56.80.
RoboTwin 2.0: 94.5% average success rate.
RoboCasa365: 49.5% overall success rate without additional embodied pretraining.
Best combined performance across all three benchmarks.
Shirt folding: after single-task post-training, two rounds of deployment-driven improvement raise success rate from 13.20% to 55.56%.