From learning individual skills to composing them into a basketball-playing agent via hierarchical RL -- introducing SkillMimic, Learning Reusable Basketball Skills from Demonstrations 🌐: 📜: 🧑🏻‍💻: Work led by Yinhuai , Qihan Zhao, and Runyi Yu.

SkillMimic leverages human demonstration data extracted from video to learn specific basketball skills like dribbling and layup

Skills are learned via contact-graph-powered HOI imitation learning

Simulated humanoid sports, rise up!

The paper proposes a novel approach called SkillMimic that enables physically simulated humanoids to learn a variety of basketball skills from human-object demonstrations. The key idea is to define skills as collections of Human-Object Interaction (HOI) state transitions and then use imitation learning to train a single policy that can acquire diverse skills. SkillMimic can effectively learn diverse basketball skills, including shooting, layups, and dribbling, within a single policy using a unified configuration. Compared to baseline methods, SkillMimic exhibits superior performance and robustness. The HLC can quickly learn complex tasks, such as continuous scoring, by reusing the skills acquired through SkillMimic. full paper:

Amazing results and amazing to see how it's evolved after PhysHOI! Looking forward to adding these data and skills into ***********. 😬

turnaround layup looks nice!

So cool!

Interesting work

Thanks!