LoopWAM: Closed-Loop Generalist Policy Improvement from Deployment Rollouts
登顶RoboChallenge Table30！LoopWAM开源，与pi的部署自进化理念不谋而合
LoopWAM, a world action model developed by Tsinghua University, Yuanli Lingji, and Jihuan Shidai, achieves top performance on RoboChallenge Table30 V2 and other benchmarks. It leverages real-world deployment data to iteratively improve robot policies through quality-guided learning, incorporating success, suboptimal, and failure trajectories to enhance world modeling and action generation. The model uses a video-action dual-branch architecture and quality steering to guide policy improvements, achieving significant performance gains across multiple tasks.
LoopWAM demonstrates a closed-loop approach to improving robot policies through real-world deployment data, integrating success, suboptimal, and failure experiences to enhance world modeling and action generation. This method achieves significant performance gains across multiple robotic tasks and benchmarks, including RoboChallenge Table30 V2 and RoboTwin 2.0.
Source: Machine Heart — Robotics on WeChat · Read original article ↗
English translation unavailable. Choose the original article to read it.
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
Reported success rate
60%
Reported success rate
100%
Reported success rate
90%
Reported success rate
80%
Reported success rate
80%
process score
92
process score
92.5
process score
89
- paper: not_reported
- code: not_reported
- dataset: not_reported
- code: Available Artifact link Open source S2
- paper: Available Artifact link Open source S2
- code: Available Artifact link Open source S2
What remains unknown
Not established in the collected evidence: Environment, Control, Data origin.
Reported performance applies to the described task. It does not establish general autonomy or deployment readiness.
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
在本次评测中，LoopWAM在11项任务上的成功率达到 60% 及以上，9项任务过程分达到80分及以上。
Open source S10
几组具体成绩，让模型能力更加直观： 整理水果、物品分类、取出绿色积木：成功率均为 100% 。
Open source S11
用小勺舀取物体：成功率 90% ，过程分 92 。
Open source S12
按下按钮：成功率 80% ，过程分 92.5 。
Open source S13
堆叠碗具：成功率 80% ，过程分 89 。
Open source S14
项目主页： https://loop-wam.github.io 论文原文： https://arxiv.org/abs/2610.18873 开源代码： https://github.com/loop-wam/LoopWAM 近日，由 清华大学、原力灵机与 极寰时代 联合研发的世界动作模型LoopWAM，登顶真机评测榜单RoboChallenge Table30 V2，成功率与过程分双双位列榜首，成绩超过VLA-DM0.5等同期方案。
Open source S2
Source：Machine Heart — Robotics on WeChat · mp.weixin.qq.com