Humanoid Goalkeeper

Learning from Position Conditioned Task-Motion Constraints

The University of Hong Kong, Shanghai AI Laboratory

Junli Ren^*, Junfeng Long^*, Tao Huang, Huayi Wang, Zirui Wang, Feiyu Jia, Wentao Zhang, Jingbo Wang^†, Ping Luo^†, Jiangmiao Pang^†

Paper

arXiv

Video Code

Humanoid Goalkeeper learns a single end-to-end RL policy, executing agile, human-like motions to intercept flying balls, as well as performing tasks such as escaping a ball using jump and squat motions.

Wide Ranges

Penalty Kicks

Continuous Performance

Esacpe Balls

Onboard Camera (without mocap)

Method

Humanoid Goalkeeper learns a single-stage, end-to-end reinforcement learning (RL) policy conditioned on ball position observations, advancing prior work by enabling:

Wide-range, whole-body, and autonomous interactions with highly dynamic objects.
Hardware-feasible deployment across different perception modalities, including mocap and onboard camera.
Generalizable skills including goalkeeping and jump/squat escaping balls.

Acknowledgements

This work was conducted during the author`s internship at the Embodied AI Center of Shanghai AI Laboratory .

We would like to thank the Photonics and Electronics Center at Shanghai AI Laboratory for supporting the motion-capture system and providing access to the experimental field.

We are especially grateful to Yuman Gao for sharing the codebase for flying ball interception using an onboard camera, adapted from his previous work Quadrupedal Robot Teams.