Two humanoid robots just played each other at table tennis, with no remote and no ball-feeder

Two humanoid robots from the Hong Kong University MMLab and KAI Robotics team played a complete 11-point table-tennis match against each other last week, with no remote control and no human feeding the ball. The team behind the demo calls it the first fully autonomous humanoid-vs-humanoid game to regulation length. The match itself is the surface; the constraint that produced it is the story.

The constraint is the rulebook. The 2nd World Humanoid Robot Games opens August 22 at the National Speed Skating Arena, known as the "Ice Ribbon," in Beijing, and runs through August 26. More than 2,000 robots are set to compete and more than 1,000 are slated for the opening ceremony lineup. Of the 26-plus events on the program, only two, ping-pong and free combat, require full autonomy. No teleoperation, no offboard ball-feeder. A rally that is just a rally is a forfeit.

That is why a sponsor-team video suddenly reads as a category milestone. The team told QbitAI that it entered the World Games precisely because the autonomy rule forced the test. "The hardest events are the ones where the system has to see, decide and move in real time, on its own," Luo said. Ping-pong is one of them.

The system tying the demo together is SMASH 2.0, a closed-loop control stack the lab has been building since October 2025. The pipeline runs from egocentric vision through ball-trajectory prediction into action planning and finally whole-body control on the robot. SMASH is not a single end-to end large model. It is a chain of specialized modules, each with its own benchmark, and the team is explicit that this modular structure is what lets them iterate.

The roadmap is concrete. SMASH 1.0 handled fixed-zone returns with a human feeding the ball. SMASH 2.0, the version on the court last week, expanded ball coverage to short and long returns and added autonomous serving. SMASH 3.0, the team's stated next target, is spin. "Right now we treat the spin as disturbance," KAI motion-control lead Li Yinghui said in the same interview. "We want to model it directly."

The honest limits of the demo are the second story. Strategies in the SMASH 2.0 match were pre-set: fast, slow, or placement. Opponent-conditioned learning is logged, not yet used at match speed. Training data is mostly human. Coaches in mocap suits produced roughly one to two months of motion data at four to eight hours a day before any robot work began. The team is candid that human pros still dominate on pace and tactics, which is why the Games also serve as a public falsifier for the lab's claims.

The research artifact exists independently of the demo. The SMASH paper sits on arXiv, and the project page documents the egocentric-vision setup and the full-body controller. The point of writing it up as a paper rather than a press release, the team says, is to make the closed-loop architecture rather than the rally count the unit of progress.

Beijing's municipal government has been framing the Games as a benchmark event for embodied AI, and the official release treats the autonomy rule as a deliberate test. Independent industry coverage from NE Times notes that the HKU approach skips the motion-capture-heavy training path some rivals have taken, betting that a perception-to-action loop built on real match data will scale better once opponent-conditioned learning is in place.

What changes between now and the closing ceremony on August 26 is whether the closed-loop stack survives contact with the rest of the field. SMASH 3.0, the spin-handling version, is the next dated milestone on the lab's public roadmap. The Games, by rule, will show whether other teams have matched the autonomy bar, and whether "autonomous" at competition speed means what SMASH 2.0 demonstrated in the lab.