Although videos of humanoid robots playing table tennis with humans are no longer rare (in these instances, the human serves the ball, and the robot returns it), the appearance of two robots playing completely autonomously without remote control or ball feeding by an operator is something new. Both robots perform all actions—serving, receiving, attacking, and defending—until the score reaches 11 points.
This match will take place at the National Oval Rink, Ice Ribbon, where the second World Robotics Tournament will be held from August 22 to 26, featuring over 2000 robots.
The participants were the HKU-KAI team, formed by researchers from the University of Hong Kong and the KAI research group of Chaowei Power. Their robots will also participate in several table tennis master matches. Before their debut, the two humanoid robots completed an entire autonomous game, following the rule of 11 points, marking the first complete autonomous table tennis match involving humanoids in history.
Players in red and blue caps exchanged shots, and the score changed back and forth. Although the game was calmer than professional level, one player served while the other received.
Behind the robots is the SMASH 2.0 system, which is a closed loop connecting perception, trajectory prediction, motion planning, and whole-body control. To conduct the full match, the team updated SMASH from version 1.0 to 2.0. This update allowed for an expanded range of shots from fixed incoming balls to short and long ones, and it added autonomous serving, which became indispensable since the robots must change sides to serve.
Team leader Lo Ping, Associate Dean of the Faculty of Computer Science and Data Science at HKU and head of MMLab HKU, noted that when two robots play, both strive to win, which requires active attack. He emphasized that a full 11-point game means not just 11 rallies, but a completed match determined according to table tennis rules.
Both robots use the same SMASH system. Although strategies can be pre-set (e.g., faster, slower, or with different landing spots), real-time learning against an opponent has not yet been achieved. Data gathered during the robot matches is saved for reinforcement learning on real robots, accumulating rallies that can contribute to self-development, although the training data is mainly obtained from humans. The team collected data over one to two months, four to eight hours daily, using motion capture trainers, and is now trying to transition to markerless data collection, sacrificing some data quality for much greater scale.
Li Yinghui, a postdoc at HKU and head of motion control algorithms at Chaowei Power, stated that table tennis and kickboxing are among the few sports requiring fully autonomous robots without external control, and table tennis has become the first formal competitive event for humanoid robots. The robot must evaluate ball trajectories, make decisions, and strike in a closed loop, demonstrating the transition of robots from executing predefined actions to real-time actions based on environmental changes. The team began working on table tennis around October 2025, and the pace of iterations in competitions has accelerated; the expanded range of shots in SMASH 2.0 was achieved through simulation training using human motion data.
From August 16 to 17, a competition on guest services was held at the Beijing Wuzhou Hotel as part of the second World Humanoid Robot Games. The contest leader, Li Ziye, noted that each round is limited to 30 minutes, during which robots must complete tasks such as baggage handling, restocking room supplies, and room cleaning. These tests assess the robots' ability for mobile weight transfer, balancing when pushed, dexterous grasping, real-environment navigation perception, and planning complex multi-stage tasks.
The goal of this competition is to accelerate the transition from laboratory development to commercial implementation. The contest includes four main challenges: unstructured real hotel environments requiring high navigation and obstacle avoidance; soft items like bedding and towels that change shape and require flexible manipulation; pushing loaded trolleys with difficult posture balance; and a long chain of tasks under time constraints, which questions the overall stability of the machine.
Zhang Yizhou from the Renmin University of China-Beijing Humanoid Alliance-iYZhou team pointed out that the venue itself was the most difficult part: real scenes differ greatly from simulated ones, and narrow corridors test the robots' maneuverability when moving forward and backward while handling luggage.
In the baggage handling task, robots must place two suitcases and two bags on a transport trolley, drive it to the specified room, unload all the luggage, and then return the trolley. For restocking supplies, robots retrieve one towel, two pairs of slippers, and two bottles of water, hang the towel straight, place the bottles on a corner cabinet, and arrange the slippers under the bedside table. During cleaning, robots smooth the bed and place two dirty towels and one piece of trash into recycling bins.
Li Ziye emphasized that timely completion of the task does not guarantee high scores; the neatness of item placement is important. Fully autonomous mode earns 1.0 point, whereas remote control mode is only rated at 0.5.
The hospitality competition is one of the events in the Games conducted in real conditions. The Games select worker robots based on real needs across nine different scenarios: home cleaning, hotel service, industrial production, emergency rescue operations, library management, retail, office service, garden inspection, and vehicle charging, developing complex tasks that test eye-brain-hand coordination.
Li Ziye added that the hospitality competition is based on real industrial needs: after surveying sites, the team identified pain points in highly standardized repetitive work, labor shortages, rising personnel costs, and high staff turnover; such work corresponds to the current stage of humanoid robot development.
Wang Fan, director of the hospitality competition and head of the Asia-Pacific office of Beichen Group, stated that as humanoid robots improve in mechanical and process services, the efficiency of hotel operations will increase. Currently, hotels already use robots for room delivery, smart cabinets, and AI-based voice stations, and plan to introduce robots for baggage handling, guest interaction, and coffee preparation, promoting a robot-oriented hotel strategy while providing more real scenarios for robot training.
