Hugging Face · Coming soon
HumanWAM-Vanilla learns from egocentric video, language, and hand motion, then transfers that experience through robot post-training.
Video + language + hand motion
50 RoboTwin tasks
50 clean demonstrations per task
Clean · Scene-OOD · Random
More egocentric pretraining improves transfer to unfamiliar scenes under both isolated and inverse-dynamics attention.
Evaluated after robot adaptation. Curves summarize the tested data ranges; each setting uses its best checkpoint.
Scale the range of tasks and scenes alongside hours of experience. Growing coverage produces steeper gains than adding trajectories within fixed coverage.
| Family | Hours | Clips | Tasks | Scenes |
|---|---|---|---|---|
| Scaled-diversity | 100 | 62,206 | 62 | 222 |
| 500 | 305,889 | 309 | 961 | |
| 1,000 | 589,216 | 547 | 1,442 | |
| Coverage-matched | 100 | 81,076 | 2,288 | 2,953 |
| 500 | 345,053 | 2,288 | 2,953 | |
| 1,000 | 653,567 | 2,288 | 2,953 | |
| Shared parent | 5,000 | 2,813,547 | 2,288 | 2,953 |
Coverage-matched keeps task and scene coverage fixed; it is labeled “Matched-diversity” in the plot.
Evaluate the pretrained checkpoint directly, then evaluate after RoboTwin post-training. The protocols are distinct; both measure average success across Clean and Random.
Rank 1 among successfully evaluated open-source checkpoints.
Select the figure to enlarge ↗
Human data: 11.48% → 13.94%.
No task post-training · Table 12.
Rank 1 in the reported RoboTwin post-training comparison.
Select the figure to enlarge ↗
92.70% Clean · 84.60% Random.
Refined post-training · Table 13.
No target-task post-training.
20 tasks · 10 trials per task · 200 trials per model.
Robot experience is fixed at 14,618 hours for these three configurations.
0 h human data
15,901 h human data
45,938 h human data
Loading demonstrations…
Selected qualitative recordings. The success rates above summarize repeated evaluation trials.
One pretrained policy, two visual targets.
Positioning excerpts from the pretrained policy. In B, the bottle tips after release in the source recording.
Same policy. Different visual targets.
After task post-training
A human-view goal directs the robot after task post-training.
After task post-training
After task-specific post-training, HumanWAM-Goal chains 14 desk-organization stages and performs mobile household tidying across multiple work areas on Astribot S1.
Mug placement, stationery collection, drawer storage, and notebook arrangement in one recorded sequence.
Selected qualitative recordings.
A person demonstrates an interaction; the robot then follows in the same continuous recording. The policy remains in inference throughout, with no mode switch between demonstration and execution.
Read the paper and explore the code. Model checkpoints and dataset links are coming soon.
@misc{ma2026openhumanscale,
title = {Open-HumanScale: An Open Recipe for Scaling Human
and Robot Experience for Embodied Foundation Model},
author = {Ma, Juncheng and Li, Jiaxin and Tu, Jiankai and
Liang, Bo and Lv, Guodong and Deng, Yufan and
Tang, Junzhe and Yang, Liangzhu and Bian, Chen and
Cao, Zicong and Zhu, Baohao and Bi, Jianxin and
Yan, Tengfei and Zhang, Yinyuan and Zhang, Kewei and
Chen, Xinglong and Chen, Kaiqi and Wang, Duomin and
Tang, Xiaotian and Li, Li and Xu, Yuan and
Yang, Haosen and Wang, Yuqi and Xia, Jingfei and
Huang, Eric and Dou, Zhiyang and Dong, Zhen and
Xie, Enze and Chua, Tat-Seng and
Zhou, Daquan},
year = {2026},
eprint = {2611.06817},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2611.06817}
}arXiv:2611.06817 · 2026