Researchers at Hong Kong University unveil groundbreaking AI system that teaches robots to dribble, shoot, and perform layups with human-like precision
A compact humanoid robot has achieved what scientists are calling a major breakthrough in robotics: the ability to play basketball autonomously with fluid, human-like movements. The demonstration, which has captivated viewers online, showcases a Unitree G1 robot, affectionately nicknamed “Little Potato”, executing three-step layups, dribbling with precision, and even recovering from blocked shots.
World’s First Real-World Basketball Robot
Researchers at the Hong Kong University of Science and Technology (HKUST), collaborating with Unitree Robotics, have achieved what they describe as the world’s first demonstration of a robot performing basketball actions in a real environment. The 1.3-meter-tall robot moved across the court with surprising fluidity during the lab demonstration, sinking shots and handling the ball with remarkable control.
In a viral video shared by HKUST PhD student Yinhuai Wang, the robot showcases its athletic capabilities while interacting with a human player. Wang attempted to block the robot’s shot, swatting away a layup attempt, yet the G1 quickly regained its balance and continued playing, demonstrating the stability of its learned movements.
Revolutionary AI Training System Behind the Achievement
The robot’s basketball prowess stems from an innovative AI framework called SkillMimic, developed by Wang and his team. The system captures human demonstrations through video and motion-tracking suits, then refines those movements through thousands of simulated drills in virtual training environments.
SkillMimic, accepted as a CVPR 2025 Highlight, introduced a unified imitation-reward mechanism and hierarchical learning framework capable of teaching robots diverse human-object interaction skills. The approach enables a single policy to learn multiple basketball skills simultaneously, allowing smooth transitions between different movements even when those transitions weren’t explicitly demonstrated in the training data.
The research has evolved to SkillMimic-V2, accepted to SIGGRAPH 2025, which tackles the challenge of learning from sparse, noisy, and imperfect demonstration data. This advancement represents a significant step forward in teaching robots complex interactive skills without requiring perfectly clean training data.
Technical Innovation: From Simulation to Reality
The robot trained using approximately 4.5 billion simulated samples to master seven basic basketball moves, including various types of dribbling, layups, shooting, and ball pickup. A high-level controller then composes these learned skills to accomplish more complex sequences, such as consecutive scoring.
The training methodology involves three key innovations:
- Stitched Trajectory Graph (STG): This technique identifies similar body positions across different skills and connects them, creating new transition paths the robot never directly observed
- State Transition Field (STF): The system trains the robot by starting from varied positions each time, including partially masked intermediate states
- Unified Learning Framework: Unlike previous methods requiring custom rewards for each skill, SkillMimic learns all basketball actions through the same configuration with identical hyperparameters
The Unitree G1: Budget-Friendly Athletic Powerhouse
The Unitree G1, starting at $16,000, has become a default choice for elite research labs due to its impressive capabilities and accessibility. With over 1,000 units shipped as of 2025, it’s currently the best-selling humanoid robot on the market.
The robot’s specifications include:
- Height: 1.32 meters
- Weight: 35 kilograms
- Up to 43 degrees of freedom in advanced configurations
- Movement speeds exceeding 2 meters per second
- Advanced sensor array including 3D LiDAR and depth cameras
Beyond basketball, the G1 has demonstrated remarkable versatility. It has achieved world records including a 1.4-meter long jump, performed side flips and kip-ups, and achieved a 90% success rate in medical task simulations. The robot has even competed in boxing matches against other humanoids.
Rising Star: Meet the Researcher Behind the Breakthrough
Yinhuai Wang, a second-year PhD student at HKUST studying under Professor Ping Tan, is the key researcher behind PhysHOI, SkillMimic, and SkillMimic-V2. His previous experience includes studies at Peking University and internships at IDEA Research, Unitree, and the Shanghai AI Lab. Online communities have dubbed him “the world’s No.1 basketball researcher” following the viral demonstration.
Wang’s work represents rapid acceleration in humanoid robotics capabilities, from early simulations in 2023 to real-world basketball demonstrations in just two years.
What This Means for the Future of Robotics
The basketball demonstration represents more than athletic showmanship. The underlying technology has significant implications for:
- Warehouse automation: Human-like mobility for navigating complex environments
- Manufacturing: Precise object manipulation in dynamic settings
- Healthcare: Robots that can assist with physically demanding tasks
- Search and rescue: Machines capable of operating in challenging terrain
The ability to learn complex interactive skills from imperfect demonstrations addresses one of robotics’ longstanding challenges. Traditional approaches required extensive programming for each specific task, but SkillMimic’s data-driven methodology allows robots to acquire diverse skills more efficiently.
The Broader Context: China’s Humanoid Robot Push
This achievement arrives amid China’s ambitious push to dominate the humanoid robotics market. The country hosted the world’s first-ever humanoid games in August 2025, featuring over 500 robots from 280 teams across 16 countries competing in 26 sports.
Unitree has established itself as a leader in this competitive landscape, with its robots performing Chinese folk dances at the Spring Festival Gala and competing in endurance challenges. A rival AgiBot robot recently entered the Guinness World Records for walking 106 kilometers over four days, showcasing the rapid pace of innovation in the field.
Looking Ahead: Next Steps for Robot Athletes
Despite impressive progress, researchers acknowledge remaining challenges. The team is working on teaching single policies to handle many different objects, learning from smaller datasets, and enabling robots to perform these skills in unstructured outdoor environments, including full basketball games.
The technology’s evolution from controlled laboratory demonstrations to unpredictable real-world scenarios represents the next frontier. As hardware continues to improve and AI training methods become more sophisticated, the gap between simulation and reality continues to narrow.
From Laboratory Demo to Real-World Revolution
The basketball-playing Unitree G1 represents a watershed moment in humanoid robotics, demonstrating that machines can now learn complex interactive skills that mirror human athletic performance. While we’re not yet ready for robot basketball leagues, this breakthrough opens possibilities that extend far beyond the court, from revolutionizing warehouse operations to enabling robots to work alongside humans in countless industries.
As research accelerates and capabilities expand, the line between science fiction and reality continues to blur.
See how humanoids athletes currently stack up against humans in 2025
