Table tennis doubles has a “counterintuitive” rule: the two players on the same team must alternate hitting the ball. Even if the ball flies right at you, if it’s not your turn, you must leave that shot to your teammate.

For robots that can already perform singles rallies, this rule brings new challenges: they must hit accurately, move quickly, and yield space in time within the limited area of their side of the table to avoid blocking teammates or colliding.

Recently, Archon Robotics officially released PingpongDuo, the first humanoid robot table tennis doubles system. Through joint positioning planning and whole-body motion control, two Unitree G1 robots are able to alternate hitting, yielding space, and returning to position — offering an implementation path for collaborative humanoid robot motion. The related paper, "PingpongDuo: Agile Humanoid Table Tennis Doubles with Safe Team Coordination," has been released as an arXiv preprint.

In this research, CHINGMU provided high-precision optical motion capture equipment, supporting the research team in human motion data acquisition, table tennis ball and paddle tracking, simulation parameter calibration, and real-robot motion state measurement — providing data and measurement basis for robot motion learning and experimental validation.

01 | From Cooperative Control to Real-World Measurement: How Do Robots Master Every Shot?

During doubles play, robots must plan their own positioning and that of their teammates, while seamlessly linking movement and hitting. The research team combined dual-robot cooperative planning, locomotion skills, and hitting control, and used real-world measurement data to calibrate the ball's flight, bounce, and contact parameters, bringing simulation training closer to real-world motion scenarios.

Experimental results show that in simulation, PingpongDuo improved the return success rate by 13.2% compared to existing state-of-the-art baseline methods. In real-robot experiments, a single robot completed yielding and returning movements with a total travel distance of 2.15 meters within 2.4 seconds, reaching a peak velocity of 2.23 m/s. The two humanoid robots achieved over 10 consecutive alternating hits, validating the system's feasibility in dual-robot coordination and dynamic motion control.

Behind these capabilities is the synergy of algorithms, data, and real-world measurement.

The full-body human motion data provided by CHINGMU serves as a reference for robot locomotion skill learning. The tracking of the ball and paddle provides a measurement basis for ball trajectory analysis and hitting interaction modeling. Physical simulation parameter calibration and real-robot motion state measurement help the research team analyze the sim-to-real gap, providing support for control strategy optimization and experimental validation.

Letting motion be precisely recorded, letting training be grounded in real data, and letting model capabilities be validated on real robots.

02 | From a Single Table to More Sports Scenarios

From a single robot swinging in rallies to two robots coordinating alternately; from the table tennis table to tennis courts, basketball courts, and go-karts — CHINGMU is continuously expanding the application of motion capture technology in robot sports. The significance of expanding these scenarios goes beyond adding a few demos; it lies in transforming every sports practice into learnable, verifiable, and reusable data resources.

The collaboration around PingpongDuo is an important practice in this direction, jointly explored by CHINGMU and Archon Robotics. Looking ahead, focusing on the R&D needs of Embodied AI models, both parties will further deepen cooperation in multi-scenario motion data acquisition, high-precision ground truth data construction, model training and validation, and robot performance evaluation — enabling real-world data to more effectively serve model R&D and real-robot experiments.

The motor capabilities of Embodied AI will not be mastered on a single table; they require continuous training and validation across diverse, real-world, and quantifiable scenarios. CHINGMU will continue to work with more ecosystem partners to promote the mutual reinforcement of high-quality motion data accumulation and model capability iteration — helping robots advance from single-agent skill breakthroughs to multi-robot coordination, and from scenario validation to continuous evolution.

Let every ball hit by a robot be more than just a brilliant rally — let it become the data starting point for the continuous learning and optimization of the next robot, the next model, and the next collaboration.