Reproduce “Learning Agile Striker Skills for Humanoid Soccer Robots from Noisy Sensory Input” on OpenResearch

About this paper

Zifan Xu, Myoungkyu Seo, Dongmyeong Lee, Hao Fu, Jiaheng Hu, Jiaxun Cui, Yuqian Jiang, Zhihan Wang and 3 more

Researchers from The University of Texas at Austin developed a four-stage reinforcement learning pipeline enabling a Booster T1 humanoid robot to acquire agile, robust, and continuous ball-kicking skills despite noisy sensory input. The method demonstrates a 79.5% success rate in simulation and 66.7% on the physical robot, achieving zero-shot transfer and improving energy efficiency.

Key insights

• Realistic modeling of perception noise, including velocity-dependent magnitude, asynchronous updates, and occlusions, is crucial for successful sim-to-real transfer of visuomotor whole-body control for humanoids.
• Constrained reinforcement learning (N-P3O) effectively refines learned policies, reducing energy consumption and mitigating unsafe or jittery movements by balancing task rewards with regularization costs, without sacrificing overall performance.
• A multi-stage teacher-student training framework, combined with online adaptation, enables humanoids to learn complex, dynamic skills that generalize to imperfect real-world sensory inputs and maintain high performance.

Summary by alphaXiv.