A team of researchers at the University of California, Berkeley, has developed a new system that enables a humanoid robot to play table tennis with a level of agility comparable to that of a human. The Unitree G1 robot — powered by HITTER (Humanoid Table Tennis Robot) — has been demonstrated in a video sustaining rallies of over 100 consecutive shots while playing with a human opponent. The now viral video was shared on X by Zhi Su, the leader of the project.
While humanoid robots have become quite sufficient at activities like walking and kicking, they have struggled with activities that require rapid interaction with fast-moving objects in dynamic environments. Table tennis is a good example of such activity, with ball speeds exceeding 5 meters per second. This mean the robot has to observe, predict, and react in the fraction of a second.
To solve this challenge, the researchers created a hierarchical framework that simulates how humans play the game. A high-level, model-based planner functions as the robot's brain, using 9 cameras to track the ball and make decisions on the action to take. A low-level controller then translates the planner's actions into coordinated body movements.
The researchers say they trained the robot's control system using human motion data, so its movements can appear more fluid and human-like. This allows the fully autonomous robot to hold and swing its paddle with fluidity, even performs side-steps, and balance itself with the freehand just like a human would. The system, which was tested on a Unitree G1 robot, provides a blueprint for the development of more such systems and robots that are even more capable.











