The Unitree G1 humanoid, nicknamed “Little Potato,” demonstrated advanced basketball skills, including dribbling, passing, and a clean three-step layup, at Hong Kong University of Science and Technology.

Its abilities come from SkillMimic, an AI framework that learns from imperfect human motion data, creating smooth transitions and robust skills through simulation.

SkillMimic-V2 uses techniques like Stitched Trajectory Graphs, State Transition Fields, and Adaptive Trajectory Sampling to help the robot recover from mistakes and execute complex sequences reliably.

The system was trained on 4.5 billion simulated samples across seven basic moves, producing reusable interaction skills that outperform traditional learning methods, Interesting Engineering has reported.

Researchers aim to expand the framework to handle more objects and eventually enable real-world humanoid robots to play full basketball games.