Humanoid robots team up to play ping-pong doubles

Watching a humanoid robot play table tennis solo is cool, but having two of them share one side of the table is a whole different ballgame. In doubles, you don't just track the ball—you have to get out of your partner's way immediately after taking a shot.

Researchers from the University of Hong Kong, Shanghai Innovation Institute, Archon Robotics, and NVIDIA introduced PingpongDuo. It is a control framework that allows two Unitree G1 humanoid robots to play table tennis doubles, sustaining rallies of over 10 consecutive alternating strikes in real-world tests.

Clearing out at over two meters per second

The real challenge in table tennis doubles is the cramped space behind the table. The ball flies fast, the alternating-strike rule is rigid, and a missed step means two expensive robots crash into each other.

PingpongDuo tackles this with a hierarchical setup. At the team level, a high-level planner calculates target positions for both robots while a risk estimator checks for hand, foot, and torso collisions. If a collision looks likely, a safety filter steps in to adjust paths, always giving priority to whichever robot needs to hit the incoming ball.

Down at the robot level, footwork and racket swings are trained as separate AI policies. The movement policy was trained on human doubles match footage, incorporating around 200 footwork sequences across various movement distances.

During live trials, a Unitree G1 robot completed a 2.15-meter clearance-and-recovery maneuver in just 2.4 seconds, reaching a peak speed of 2.23 meters per second. Researchers reported clearance positioning errors of roughly 2 centimeters. Across simulations, the framework boosted ball-return success by 13.2% compared to previous state-of-the-art baselines.

There is still a catch. Right now, the setup depends on an external motion-capture system to track the ball and the robots, and the algorithm does not factor in ball spin. The researchers plan to tackle onboard perception, spin estimation, and tactical strategies in future work.

Why it matters

We think the big deal here has little to do with ping-pong trophies and everything to do with multi-robot spatial coordination.

Most humanoid demonstrations show one robot doing one task in isolation. Put two of them in a tight kitchen, a crowded factory aisle, or a collapsed building, and they usually freeze up or slow to a crawl just to avoid bumping into each other.

PingpongDuo pushes cooperative dynamics to an extreme: split-second timing, sub-second role swapping, and rapid lateral sprints within inches of each other. Showing that lightweight models can juggle team-level collision filters with dynamic footwork is a massive technical win.

It sets a clear benchmark for how agile physical agents will share confined workspaces in the near future. While the reliance on external optical tracking proves it is still confined to the lab, the concept of fluent robotic teamwork just took a giant leap forward.