Li Fei-Fei's Team Builds a World Action Model: Switching the Base, Robot Success Rate Soars Directly
李飞飞团队造了个世界动作模型：换个底座，机器人成功率直接飙升
Stanford's Li Fei-Fei team introduces OpenWAM, an open framework for world-action models (WAM) that standardizes training and evaluation. The framework uses causal video pre-training, modular components like IDM and FDM, and a shared MoT architecture to improve task success rates across diverse robotic tasks. The paper highlights the importance of sim-to-real transfer and the value of diverse, counterfactual data in training robust action models.
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
training days
14
number of segments
32,000
Reported success rate
92.1%
- paper: not_reported
- code: not_reported
- weights: not_reported
- dataset: not_reported
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.
同一副骨架，四种「先想还是先动」 OpenWAM 从阿里开源的视频模型 Wan2.2-5B 出发，先在约 334 万条机器人与人类交互视频上继续预训练，名义时长约 1.46 万小时，全程不用动作标签，在 32 块 B200 上训练了 14 天。
Open source S11
为此团队构建了 LIBERO-Long-CF 数据集 ：恢复示范中的仿真器状态，执行被改动过的动作，比如停止、反向、加噪声、改变夹爪开合时机等，共得到 32000 个片段、约 410 万条控制记录，是原始示范的 29.7 倍，其中 72.3% 的片段改变了物体的摆放状态。
Open source S14
实验：底座很能打，翻译器能搬家 在 LIBERO 四个子集上，四种交互程序的成功率都很高，VTA 平均达 98.6%，略高于论文引用的 LingBot-VA（98.5%）和 π0.5（96.9%）等已报告结果。
Open source S15
真机上，团队用两台 Franka FR3 机械臂测试了烤面包、还原 2×2 魔方最后一层和按颜色分拣杯子，VTA 和 Joint 平均成功率分别为 92.1% 和 91.9%。
Open source S18
消融实验显示了底座的分量：其他条件不变，用原版 Wan2.2 初始化时 VTA 在 LIBERO-Long 上只有 68.4%， 换成机器人视频预训练的因果底座后升至 97.8%，提升 29.4 个百分点 。
Open source S19
从同一状态出发执行 16 种不同动作，让模型判断预测画面对应哪个真实结果，随机猜测约为 6%，只用示范训练的 FDM 为 21.1%，加入反事实数据后升至 71.3%，模型终于能分清「这样推」和「那样推」的区别。
Open source S22
Source：Machine Heart — Robotics on WeChat · mp.weixin.qq.com