A humanoid robot that plays ping pong exceeds average

To what extent can these machines compete with our own species.

Researchers at the University of California at Berkley showed the HITTER, a humanoid robot trained to play table tennis, which has already trained more than 100 times in a row against human opponents with reflexes and movements reminiscent of a beginner in the sport.

The video released by the institution, the HITTER, shows a certain naturalness in the movements, in each play, the right arm executes the blows with the racket, while the left hand extends to maintain balance, reproducing characteristic gestures of real players.

In live tests, the robot performed a sequence of 106 conceptual returns against humans, a performance that would surpass most amateurs.

The explanation for this feat is in the double system design added to the Robot, on top, a high-level planner works as the brain of the machine, using external cameras to track the ball in real time, predicts where the ball is going to land and calculates the exact position, speed and ideal time for the perfect return.

But thinking is not enough.

A low-level complementary system acts as the body, transforming those commands into coordinated arm and leg movements. This combination allows the robot to move laterally, turn and swing the racket fluidly, responding in less than a second to balls traveling up to 5 m per second.

The challenge of teaching a robot to play table tennis is gigantic.

The sport requires second-hand decisions and violations, adaptation to unpredictable plays and precise coordination. To overcome these barriers, the team combined two worlds: model-based planning, responsible for predicting trajectories and choosing actions, and reinforcement learning, which adjusts movements by trial and error.

By also integrating human movement data, the HITTER began to move more intuitively, even imitating the position of the feet and the rotation of the body, details that make it incredibly realistic.

The robustness of the system was tested on a Unitree platform, the G1, playing not only against humans, but even against another robot, this test represents a significant step in building machines capable of interacting with the world in increasingly human ways.

By combining hierarchical planning, reinforcement learning and observation of real data, the project paves the way for more agile, intuitive and capable robots. And perhaps the big question is not whether they will beat amateur players, but to what extent these machines will be able to compete with our own species in activities that we always think of as exclusively human.