A Robot That Plays Table Tennis Better Than You? 🤯

Dynamic tasks like table tennis push humanoid robots to their limits. Tracking a ball flying at over 5 m/s, predicting its trajectory, and coordinating full-body movements in fractions of a second isn’t easy for humans — let alone robots.

Researchers at UC Berkeley have developed HITTER (Humanoid Table Tennis Robot) to take on this challenge. By combining a physics-based planner with a reinforcement learning whole-body controller trained on human motions, HITTER can sustain rallies of 100+ consecutive shots with over 92% accuracy — and it even chooses forehand vs. backhand strokes like a human player.

How HITTER Works

The innovation lies in its hierarchical design:

• Integrated paddle end-effector designed for consistent, repeatable strikes

• Ball tracking via motion capture to provide real-time trajectory and ball position data

• Model-based planner predicting strike time, racket position/velocity, and landing target using physics-informed models

• Learning-based whole-body controller (WBC) trained with human reference motions, enabling natural swings while keeping balance

• Planner + learning integration: high-level strike goals are translated into low-level joint actions for agile, human-like rallies

This blend of planning and learning doesn’t just make the robot good at table tennis — it shows how humanoids can handle fast, reactive tasks that require real-time perception and coordination.  Congrats to the team on this Zhi Su, Bike Zhang, Nima Rahmanain, Yuman Gao, Qiayang Liao, Caitlin Regan, Koushil Sreenath, Shankar Sastry, keep going 🦾🦾🦾

Learn More

HITTER Site: https://humanoid-table-tennis.github.io/
Written by Kyle Hulse
Follow on Linkedin for more drones, robotics, and ai updates!