Humanoid robots play a full 11-point table tennis match with no remote control

Two humanoid robots played a complete 11-point table tennis match with no remote control and no human feeding balls, trading serves and rallies until a winner emerged. The demonstration was delivered by the HKU-Chaowei team, a collaboration between the University of Hong Kong and the KAI research team of Chaowei Power, ahead of the upcoming World Humanoid Robot Games.

The match previews the second World Humanoid Robot Games, which will be held August 22-26 at the National Speed Skating Oval "Ice Ribbon" in Beijing. More than 2,000 robots are expected to compete, with over 1,000 appearing together at the opening ceremony. Table tennis is one of the official competitive events, and the team's robots will also face off against several table tennis stars during the opening.

The autonomous rallying is powered by the team's self-developed SMASH system. Rather than a single end-to-end large model, it is a closed-loop pipeline that ties together visual perception, trajectory prediction, motion planning, and whole-body control. For this match, the team upgraded SMASH from 1.0 to 2.0: the hitting range now covers short and long balls instead of relatively fixed trajectories, and autonomous serving was added — a requirement under 11-point rules, since robots must alternate serving to complete a full game.

Key team members include Luo Ping, associate dean of HKU's Faculty of Computing and Data Science, head of MMLab HKU and co-founder of Chaowei Power, and Li Yinghui, a postdoctoral researcher at HKU and head of motion-control algorithms at Chaowei Power. The team began working on table tennis around October 2025; the competition is one milestone in that process, though it has clearly accelerated iteration.

Robot-versus-robot play is far harder than human-versus-robot play. Luo explains that humans instinctively feed balls to make rallies easier, whereas two robots facing each other both aim to win: they must attack actively while keeping the rally going. A full 11-point game is also not the same as 11 consecutive rallies — it means completing an entire game under table tennis rules and deciding a winner.

On data, the team spent one to two months collecting human motion data, four to eight hours per day, with coaches wearing motion-capture suits to produce high-quality recordings. Data from robot-versus-robot rallies is saved for later real-machine reinforcement learning. However, the team concedes the robots cannot yet learn in real time during a match; online learning would be needed for them to adapt tactics to a specific opponent.

Competition rules require all teams to use the AgiBot Yuanzheng A3 humanoid robot to eliminate hardware differences. The team says SMASH has already been deployed and tested on Unitree's G1, that algorithm generality is a long-term priority, and that it will keep adapting SMASH to different platforms while deepening cooperation with AgiBot.

Why table tennis? Li says it is the national sport, with broad appeal and strong interactivity, and a demanding proving ground for high-speed ball perception, real-time decision-making, and whole-body control. In her view, it represents robots moving from executing pre-set actions to acting in real time based on changing environments.

The team says SMASH 3.0 will focus on spin balls, the hardest hurdle on the way to professional-level play. After marathons validated stability and endurance, table tennis now validates perception, decision-making, and interaction — a controlled arena where embodied AI can stress-test its technology path. Whether humanoid robots can soon truly challenge human players is the next question to watch.