Unitree G1 Robot Plays Table Tennis with Incredible Agility

Key Takeaways

1. Scientists at UC Berkeley developed a humanoid robot, Unitree G1, capable of playing table tennis nearly as well as a human.
2. The robot uses a system called HITTER, allowing it to maintain rallies of over 100 shots against a human player.
3. Humanoid robots struggle with fast-paced sports due to the need for quick reactions to fast-moving objects.
4. The researchers created a hierarchical framework that mimics human play, using a high-level planner and a low-level controller for movement.
5. The robot’s control system was trained on human movement data, enabling smooth and lifelike actions, paving the way for more advanced robotic systems.

A group of scientists at the University of California, Berkeley, has created a new system allowing a humanoid robot to play table tennis nearly as well as a human. The Unitree G1 robot, which operates using HITTER (Humanoid Table Tennis Robot), has been showcased in a video where it managed to keep rallies going for over 100 shots while competing against a human player. This video has gained a lot of attention online, thanks to Zhi Su, the head of the project, who shared it on X.

Challenges with Fast-paced Sports

Humanoid robots have made significant progress in tasks like walking and kicking, but they often have trouble with activities that demand quick responses to fast-moving objects in unpredictable settings. Table tennis exemplifies this, with balls flying at speeds greater than 5 meters per second. This means the robot needs to watch, anticipate, and react in mere moments.

Innovative Framework for Robot Play

To tackle this issue, the researchers devised a hierarchical framework that mimics how humans engage in the game. A high-level, model-based planner acts as the robot’s brain, utilizing 9 cameras to monitor the ball and decide the best course of action. Following that, a low-level controller converts the planner’s decisions into synchronized movements of the robot’s body.

The team indicates that they trained the robot’s control system using data from human movements, which helps its actions look more smooth and lifelike. As a result, the fully autonomous robot can grip and swing its paddle fluidly, and even execute side-steps and balance itself with its free hand, just as a human would do. The system, tested on the Unitree G1 robot, offers a model for creating even more advanced robots and systems in the future.

Source:
Link