How SMASH 2.0 Completed a Fully Autonomous 11-Point Table-Tennis Game

The HKU–KAI/Chaowei team’s SMASH 2.0 demonstration was significant because it showed a complete contest rather than an isolated robotic trick. The team says two humanoid robots independently played an 11-point table-tennis game, alternating serves, responding to incoming shots and accumulating points until one robot won. The demonstration used no remote control, human ball-feeding or external cameras. 4612

The “first” wording should be read as the researchers’ claim, rather than as an independently established industry-wide finding. What is well supported by the available reporting is the scope of the demonstration: autonomous humanoid-versus-humanoid play using onboard perception and a full-game format. 24

A full 11-point game is more than 11 rallies

A rally is the exchange of one ball until a point ends. A full 11-point game requires the robots to manage the contest around those exchanges: serving, receiving, returning the ball, handling missed shots, attributing points, changing service and continuing until the rules produce a winner. Eleven successful back-and-forth exchanges would not by themselves demonstrate those capabilities. 57

That distinction matters technically. A robot can be made to return balls in a cooperative drill or respond to a human feeder without possessing the broader autonomy needed to start and complete a competitive game. Two robots playing to win must also produce offensive and defensive responses instead of simply sustaining a demonstration rally. 57

How SMASH 2.0 closes the perception-to-action loop

SMASH 2.0 is described as a modular system that links four functions:

1. Perception: Onboard, egocentric cameras detect the ball and estimate relevant robot state without relying on an external camera rig. 2
2. Trajectory prediction: The system calculates where the ball is travelling and anticipates its flight and bounce. 714
3. Motion and stroke planning: It selects how and where to hit the ball, including responses to different incoming trajectories. 714
4. Whole-body control: The robot coordinates its legs, torso, arms and racket to reach the ball while maintaining balance and executing the selected stroke. 214

Reports describe this as a millisecond-scale decision and control window: the robot must identify the incoming ball, calculate its trajectory and plan a stroke before the next playable contact. 714 The academic description of SMASH similarly emphasizes task-aligned imitation learning for strike control and onboard perception for real-time interaction. 2

What changed from SMASH 1.0

The reported upgrade from SMASH 1.0 focused on capabilities required for a real game rather than a convenient return drill.

Autonomous serving

SMASH 2.0 added autonomous serving, allowing the robots to initiate points and alternate serves according to the game format. Without that capability, a robot could participate in exchanges but could not independently run a complete match from start to finish. 58

A wider hitting range

The system expanded the robot’s reachable striking space to handle both shorter and longer incoming balls. The team said it used human-motion data in simulation to help build a broader library of striking motions and improve coverage of the robot’s workspace. 2

More than cooperative returns

A complete match also requires the robots to attack and defend. The team’s description of SMASH 2.0 includes different return types and target placements, making the system more match-oriented than one that merely returns a predictable feed. However, the available sources do not establish that the robots possess human-level tactical versatility; the stronger conclusion is that the demonstration connected autonomous serving, perception, movement and scoring into one game loop. 714

The system’s current limit: preset strategy, not online adaptation

The robots can currently be given preset approaches, such as faster or slower play or different target placements. The team says they have not yet achieved genuine online learning against a specific opponent during a live match. In practical terms, the robot cannot continually infer a rival’s preferences and revise its strategy on the fly.

The proposed next step is to preserve data from robot-versus-robot matches—including successful and unsuccessful exchanges—and use those records for later real-robot reinforcement learning. That could eventually support the team’s idea of robotic “self-evolution,” but it remains a future training objective rather than a demonstrated capability of the reported match.

How the team says the robots were trained

The team reports that its current training data are primarily human data. Human-motion recordings were used in simulation to expand the robots’ striking range, while the SMASH research describes motion-capture demonstrations and generated motion libraries as part of the system’s skill-learning approach. 2

More specific reports describe daily coaching sessions lasting roughly one to two months with motion-capture equipment, followed by interest in markerless motion collection despite possible compromises in data quality. That detail is best treated as team-reported methodology; the sources available here do not independently verify every part of the process. 35

Why table tennis is a demanding humanoid benchmark

Table tennis compresses several difficult robotics problems into a very short time. A humanoid must track a small, fast-moving ball, predict its trajectory and bounce, choose a response, move into position, coordinate multiple joints, preserve balance and make controlled contact. The opponent and the physical environment are changing continuously. 1418

That makes the sport a useful adversarial test of embodied AI. Success depends on responding to new conditions rather than replaying a fixed sequence. It is also why table tennis can be discussed alongside other formal robot-sport challenges, including combat events: both require fast reactions, physical coordination and decisions made in response to an opponent.

From demonstration to competition in Beijing

The second World Humanoid Robot Games took place at Beijing’s National Speed Skating Oval, known as the “Ice Ribbon,” from August 22 to 26, 2026. Table tennis debuted as an official competition event, with 12 teams participating. 810

The broader Games drew 2,056 robots across 666 teams and 51 disciplines, although that overall figure should not be confused with the number of table-tennis robots. The HKU SMASH team ultimately reached the table-tennis final and finished runner-up to the PKU–BAAI team, which won 2–0. 110

The SMASH robots also appeared in a human exhibition involving Olympic table-tennis champion Ding Ning. Reports identify Ding Ning as the clearly documented human opponent for the HKU system; other athletes appeared in separate demonstrations at the Games. 111920

Who built SMASH?

The project is a collaboration between University of Hong Kong researchers and the embodied-AI company KAI, also referred to in reporting as Chaowei or Chaowei Motion. The HKU SMASH team is described as a young embodied-AI research group led by researchers and largely composed of HKU doctoral students, with Professor Lo Ping identified as a supervisor and research leader. 811

The team reported spending about six months enabling the robots to play against people, followed by approximately half a month preparing two robots for autonomous play against each other. That timeline describes the reported preparation for the demonstration, not the full history of the underlying table-tennis research, whose precise start date is not established in the available sources. 11

The practical achievement of SMASH 2.0 is therefore narrower—and more useful—than the broadest headlines suggest: it demonstrated a claimed first complete autonomous humanoid-versus-humanoid table-tennis game, while also exposing the next frontier. The robots can now connect perception, prediction, movement and scoring in a competitive format; they still need reliable online opponent modeling and continual learning before they can be described as strategically adaptive players.