Two Humanoid Robots Complete an Autonomous 11-Point Table Tennis Match

The red-capped robot serves. The blue-capped robot receives. Then they trade returns until one side reaches 11 points, completing a full table tennis game without remote control or a human feeding balls. The HKU-KAI team, researchers from the University of Hong Kong and Chaowei Power’s KAI team, calls it the first complete autonomous humanoid table tennis match in history.

The performance runs on SMASH 2.0, a closed-loop system that links perception, trajectory prediction, motion planning and whole-body control. The upgrade from SMASH 1.0 expanded the robots’ hitting range from fixed incoming balls to long and short balls, while autonomous serving solved a basic problem in robot-versus-robot play: each player must serve in turn. Team leader Luo Ping said that a full 11-point game means more than completing 11 rallies; it means winning under table tennis rules.

The two robots use the same system. Their strategies can be set in advance, including shot speed and landing points, but online learning against an opponent has not yet been achieved. The team is saving robot-match data for real-robot reinforcement learning, while much of its current training data still comes from humans. Coaches wore motion-capture equipment for one to two months, training four to eight hours daily; the researchers are now testing markerless collection to gather more data, accepting some loss in quality for greater scale.

The team began working on table tennis around October 2025, and competition helped accelerate the next iteration. SMASH 2.0’s wider hitting range came from simulation training using human-motion data. For Li Yinghui, an HKU postdoctoral researcher and Chaowei Power’s motion-control algorithm lead, table tennis is one of the few events that demands fully autonomous robots: the machine must judge the ball’s path, choose a response and strike inside one continuous loop.

So what changes in practice? The robots are moving from executing preset actions toward reacting to live changes in their surroundings. The team presents the match as a test of a robot’s ability to judge the ball’s path, make decisions and strike in a closed loop. The immediate arena is the Ice Ribbon, where the pair will play at the second World Robot Games from August 22-26 and may also face table tennis masters. The longer-term payoff remains conditional: the robots still lack online learning, and their training pipeline is still largely built on human data.