Teaching a humanoid robot to play table tennis is a well-established benchmark for high-speed control. But teaching two of them to play doubles is an entirely different beast—one that’s less about a killer topspin and more about not tripping over your multi-thousand-dollar partner. Researchers from Archon Robotics, NVIDIA, The University of Hong Kong, and the Shanghai Innovation Institute have tackled this chaotic dance with a new system called PingpongDuo. In real-world experiments, two Unitree G1 humanoids successfully sustained rallies of more than 10 consecutive shots, all while gracefully (mostly) staying out of each other’s way.
The core challenge of doubles isn’t hitting the ball; it’s the frantic repositioning in a confined space. Teammates must strike in strict alternation, meaning the moment one robot hits the ball, it has to clear out for its partner with less than two seconds to spare. A moment’s hesitation risks a collision between two 35kg machines, turning a friendly match into a very expensive pile of actuators and shame. PingpongDuo addresses this with a hierarchical framework detailed in a paper available on arXiv. At the team level, a planner selects positioning targets and uses a learned “risk filter” to veto any movement plans likely to cause a friendly fire incident.
At the individual level, the robots use separate skills for footwork and striking, trained on motion data from human doubles players. The system is brutally efficient: one robot was clocked completing a 2.15-meter clearance-and-recovery maneuver in just 2.4 seconds, hitting a peak speed of 2.23 m/s. In simulation, PingpongDuo improved the ball-return success rate by 13.2% over the previous state-of-the-art singles-playing robot. You can see videos of the robots in action on the official project page.
Why is this important?
This research isn’t just a stepping stone to an all-robot Olympics. It’s a serious exploration of dynamic, safe coordination for multiple robots operating in a shared physical space. The AI framework that prevents two humanoids from smashing into each other during a fast-paced game could be adapted for complex warehouse logistics, collaborative assembly lines, or any scenario where multiple autonomous systems need to work in close quarters without causing expensive, metallic carnage. For Archon Robotics, this is a powerful demonstration of its broader “Whole-Body Intelligence” platform, proving it can manage not just one robot’s body, but a team’s.