Humanoid Policy   Human Policy

Ri-Zhao Qiu, Shiqi Yang, Xuxin Cheng, Chaitanya Chawla, Jialong Li, Tairan He, Ge Yan, David J. Yoon, Ryan Hoque, Lars Paulsen, Ge Yang, Jian Zhang, Sha Yi, Guanya Shi, Xiaolong Wang
Proceedings of The 9th Conference on Robot Learning, PMLR 305:2888-2906, 2025.

Abstract

Training manipulation policies for humanoid robots with diverse data enhances their robustness and generalization across tasks and platforms. However, learning solely from robot demonstrations is labor-intensive, requiring expensive tele-operated data collection,n which is difficult to scale. This paper investigates a more scalable data source, egocentric human demonstrations, to serve as cross-embodiment training data for robot learning. We mitigate the embodiment gap between humanoids and humans from both the data and modeling perspectives. We collect an egocentric task-oriented dataset that is directly aligned with humanoid manipulation demonstrations. We then train a human-humanoid behavior policy, which we term Human Action Transformer (HAT). The state-action space of HAT is unified for both humans and humanoid robots and can be differentiably retargeted to robot actions. Co-trained with smaller-scale robot data, HAT directly models humanoid robots and humans as different embodiments without additional supervision. We show that human data improves both the generalization and robustness of HAT with significantly better data collection efficiency.

Cite this Paper

BibTeX
@InProceedings{pmlr-v305-qiu25a, title = {Humanoid Policy   Human Policy}, author = {Qiu, Ri-Zhao and Yang, Shiqi and Cheng, Xuxin and Chawla, Chaitanya and Li, Jialong and He, Tairan and Yan, Ge and Yoon, David J. and Hoque, Ryan and Paulsen, Lars and Yang, Ge and Zhang, Jian and Yi, Sha and Shi, Guanya and Wang, Xiaolong}, booktitle = {Proceedings of The 9th Conference on Robot Learning}, pages = {2888--2906}, year = {2025}, editor = {Lim, Joseph and Song, Shuran and Park, Hae-Won}, volume = {305}, series = {Proceedings of Machine Learning Research}, month = {27--30 Sep}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v305/main/assets/qiu25a/qiu25a.pdf}, url = {https://proceedings.mlr.press/v305/qiu25a.html}, abstract = {Training manipulation policies for humanoid robots with diverse data enhances their robustness and generalization across tasks and platforms. However, learning solely from robot demonstrations is labor-intensive, requiring expensive tele-operated data collection,n which is difficult to scale. This paper investigates a more scalable data source, egocentric human demonstrations, to serve as cross-embodiment training data for robot learning. We mitigate the embodiment gap between humanoids and humans from both the data and modeling perspectives. We collect an egocentric task-oriented dataset that is directly aligned with humanoid manipulation demonstrations. We then train a human-humanoid behavior policy, which we term Human Action Transformer (HAT). The state-action space of HAT is unified for both humans and humanoid robots and can be differentiably retargeted to robot actions. Co-trained with smaller-scale robot data, HAT directly models humanoid robots and humans as different embodiments without additional supervision. We show that human data improves both the generalization and robustness of HAT with significantly better data collection efficiency.} }
Endnote
%0 Conference Paper %T Humanoid Policy   Human Policy %A Ri-Zhao Qiu %A Shiqi Yang %A Xuxin Cheng %A Chaitanya Chawla %A Jialong Li %A Tairan He %A Ge Yan %A David J. Yoon %A Ryan Hoque %A Lars Paulsen %A Ge Yang %A Jian Zhang %A Sha Yi %A Guanya Shi %A Xiaolong Wang %B Proceedings of The 9th Conference on Robot Learning %C Proceedings of Machine Learning Research %D 2025 %E Joseph Lim %E Shuran Song %E Hae-Won Park %F pmlr-v305-qiu25a %I PMLR %P 2888--2906 %U https://proceedings.mlr.press/v305/qiu25a.html %V 305 %X Training manipulation policies for humanoid robots with diverse data enhances their robustness and generalization across tasks and platforms. However, learning solely from robot demonstrations is labor-intensive, requiring expensive tele-operated data collection,n which is difficult to scale. This paper investigates a more scalable data source, egocentric human demonstrations, to serve as cross-embodiment training data for robot learning. We mitigate the embodiment gap between humanoids and humans from both the data and modeling perspectives. We collect an egocentric task-oriented dataset that is directly aligned with humanoid manipulation demonstrations. We then train a human-humanoid behavior policy, which we term Human Action Transformer (HAT). The state-action space of HAT is unified for both humans and humanoid robots and can be differentiably retargeted to robot actions. Co-trained with smaller-scale robot data, HAT directly models humanoid robots and humans as different embodiments without additional supervision. We show that human data improves both the generalization and robustness of HAT with significantly better data collection efficiency.
APA
Qiu, R., Yang, S., Cheng, X., Chawla, C., Li, J., He, T., Yan, G., Yoon, D.J., Hoque, R., Paulsen, L., Yang, G., Zhang, J., Yi, S., Shi, G. & Wang, X.. (2025). Humanoid Policy   Human Policy. Proceedings of The 9th Conference on Robot Learning, in Proceedings of Machine Learning Research 305:2888-2906 Available from https://proceedings.mlr.press/v305/qiu25a.html.

Related Material