UC Berkeley researchers just taught a general-purpose humanoid to rally at the ping-pong table, and it’s as impressive as it sounds. Their new system, called HITTER (Humanoid Table Tennis Robot), turns the Unitree G1 into a quick-thinking, quick-moving opponent capable of sustaining rallies of more than 100 consecutive shots against a human player. The team published its work on August 28, 2025, and shared videos that have since drawn wide attention, including a clip posted on X by project lead Zhi Su.
Why this matters: most humanoid robots can walk, balance, and even kick, but they tend to struggle when faced with fast, unpredictable objects. Table tennis pushes the limits of perception and control, with balls often traveling over 5 meters per second. To compete, a robot has to see the ball, predict its trajectory, decide how to respond, and move its body—all in fractions of a second.
The team’s solution is a hierarchical system that mirrors how people play. A high-level, model-based planner acts like the robot’s brain, fed by a 9‑camera vision setup that tracks the ball and chooses the next shot. Beneath that, a low-level controller translates those decisions into coordinated, full-body movements. The result is a fully autonomous robot that not only swings a paddle with convincing fluidity but also side-steps, repositions, and stabilizes with its free hand—just like a human.
Training was a key ingredient. By incorporating human motion data, the system produces smoother, more natural movements that look less like scripted robotics and more like real athletic play. According to the researchers, the framework tested on the Unitree G1 offers a blueprint for building future humanoids that can handle fast, dynamic tasks in the real world.
Key takeaways:
– A general-purpose humanoid robot, the Unitree G1, can now rally in table tennis using HITTER.
– The system combines a 9-camera vision array, a high-level model-based planner, and a low-level motion controller.
– Trained with human motion data, the robot achieves human-like fluidity, sidestepping and balancing during play.
– Demonstrated rallies exceed 100 consecutive shots against a human opponent.
– The research, published on August 28, 2025, highlights a path toward robots that can react reliably in high-speed, dynamic environments.
Beyond the ping-pong table, this kind of perception-to-action pipeline could translate to warehouse picking, collaborative manufacturing, sports training, rehabilitation, and any job that demands rapid decisions and whole-body coordination. If robots can master a rally at high speed, they’re on track to handle the unpredictability of everyday tasks with far greater agility.
