PingpongDuo: Humanoid Table Tennis Doubles with Safe Team Coordination

· Editorial Team estimated
humanoid table-tennis multi-robot-coordination motion-planning skill-learning legged-locomotion

Humanoid table tennis doubles forces two robots to strike in strict alternation while sharing a confined workspace, where fast repositioning sacrifices striking accuracy and uncoordinated movement risks collisions. PingpongDuo resolves the conflict with a hierarchical framework: at the team level a planner jointly selects positioning targets and filters risky proposals with learned risk estimates, preferentially preserving the current striker's target; at the individual level separate footwork and striking skills are trained from one another's states and guided by human motion references, including footwork from human doubles play; a racket-arm residual then refines strikes using feedback from a simulator calibrated with real ball-flight and contact measurements. In simulation it improves return success by 13.2% over the state of the art at roughly 2 cm positioning error; on hardware one robot completes a 2.15 m clearance-and-recovery cycle in 2.4 s at a 2.23 m/s peak, and two robots sustain more than 10 consecutive strikes in strict alternation.

Paper · arXiv:2610.20579

Singles table tennis is already demanding; doubles is a level harder — two robots must strike in strict alternation while sharing one confined court.

Background

Humanoid locomotion keeps advancing, but most public demonstrations remain single-robot and single-task. Doubles raises the difficulty by an order of magnitude: two robots must complete strikes in strict alternation while occupying the same near-table space, where one robot’s fast repositioning can crowd out its teammate’s stance or even cause a collision. Moving to save a shot often costs striking accuracy, while overly cautious coordination makes the robot miss the best contact point and hand the advantage to the opponent.

Core Innovation

PingpongDuo resolves this with a hierarchical framework. At the team level, a planner jointly selects positioning targets for both players and uses learned risk estimates to filter out risky proposals, preferentially preserving the current striker’s target so a strike is not interrupted by clearance. At the individual level, capability is split into separate footwork and striking skills: footwork handles agile movement and striking handles precise racket control. Both skills are guided by human motion references, including footwork collected from human doubles play, and are trained from states produced by one another to support rapid transitions. Finally, a racket-arm residual refines the strike after movement, using return feedback from a simulator calibrated with real ball-flight and contact measurements.

Results

  • In simulation, the ball-return success rate improves by 13.2% over the state-of-the-art baseline, with roughly 2 cm positioning error during clearance.
  • On hardware, a single robot completes a 2.15 m clearance-and-recovery cycle in 2.4 s, reaching a peak speed of 2.23 m/s.
  • Two robots sustain more than 10 consecutive strikes in strict alternation.

Limitations

The abstract does not report the distribution of failure modes, nor behavior on a larger court or against faster balls. Although the real-to-simulation gap is calibrated with real ball-flight and contact data, the human-doubles footwork reference is limited in coverage, and whether it transfers to more varied incoming styles is untested. Whether safe coordination for robots sharing a confined space extends beyond a structured table into unstructured settings also remains open.

Industry Implications

The contribution is not about table tennis. The question actually answered is how two robots should safely divide labor and yield space when they share one physical workspace. The same conflicts — rotation, clearance, collision avoidance — will arise in nearly every multi-robot coordination setting ahead, from assembly cells to warehouse picking. Decoupling team-level planning from individual-level skills is a reusable engineering pattern, and the fact that a meter-scale clearance completes in seconds on hardware shows that legged platforms now carry enough dynamic margin for adversarial, tightly timed multi-robot tasks. For a sector whose public face is the humanoid demo, this is a clear signal of the shift from single-robot showpieces toward coordinated work.