Motor training
The motor policy first learns to return prescribed balls to requested landing targets.
From Hitter to Player
Separating Motor Execution and Strategy
in Competitive Humanoid Table Tennis
1 Zhejiang University
2 Ocean University of China
3 Tongji University
4 Shanghai Innovation Institute
5 Tsinghua University
6 RoboParty Lab
Strategy comparison in simulation
The learned strategy wins more points against Center, Random, and Away while every selector uses the same frozen motor stack.
- vs. Center
- 88.9 ± 1.4%
- Always selects the center region.
- vs. Random
- 81.9 ± 2.0%
- Selects one of nine regions at random.
- vs. Away
- 73.1 ± 4.3%
- Selects the region farthest from the opponent.
Evaluation protocol
200 simulated points per pair per seed, eight seeds, balanced across table side and initial receiver. Win rates are mean ± sample SD across seeds. The learned strategy selects by argmax, without target jitter. The motor actor, observation normalizer, planner, and joint controller are fixed for all selectors.
Abstract
Competitive humanoid table tennis requires two distinct abilities: executing a desired return and deciding where to return the ball. We separate these two problems, with the ball landing target serving as an interface between motor control and strategy. (1) We first train a whole-body motor policy conditioned on the desired landing target, which maps the target to an executable humanoid stroke through a model-based strike planner and a learned controller. (2) We then improve the motor policy through self-play in closed-loop rallies, where the robot learns to handle incoming balls generated by another player rather than only prescribed incoming balls. At this stage, the landing targets are still sampled without considering the opponent, so the policy improves its execution without learning game strategy. (3) Finally, we freeze the entire motor stack and train a stroke-level strategic policy to choose landing targets through competitive self-play.
Motor self-play improves both landing accuracy and legal-return rate during rallies. More importantly, with the entire motor stack fixed, changing only the target-selection policy leads to large differences in game outcomes: the learned strategy wins 89%, 82%, and 73% of points against center, random, and away-from-opponent baselines, respectively. The learned strategy also changes where it sends the ball according to the opponent's position and the previous target, showing that self-play learns more than simply aiming at fixed “good” locations. On hardware, the same frozen execution stack and strategic interface support complete 11-point games against a recreational human opponent, including multi-stroke rallies with both forehand and backhand returns.
Human–robot matches
Game 02 / G1 11 : Human 5
Four games against a recreational human opponent, using the same frozen motor stack.
Target-conditioned returns
One motor policy returns the ball to different landing targets.
Landing target 4Real robot
Changing landing targetsSimulation
Three-stage training
The strategy chooses where the ball should land; the motor policy executes the return.
Motor self-play
Self-play adapts it to rally-generated balls, with targets sampled independently of opponent position.
Strategic self-play
The motor stack is then frozen while competitive self-play learns where to return.
From landing target to joint commands
The same landing target conditions the strike planner and whole-body controller.
Execution and target selection
Landing accuracy
Requested targets and achieved landing positions in simulation and on hardware.
Simulation
Requested targetActual legal landing
- Legal first landings
- 99.2%
- Within 30 cm
- 94.3%
- Landing error, mean ± SD
- 15.0 ± 8.4 cm
Hardware
Reported targetActual legal first landing
- Within 45 cm
- 69.3%
- Landing error, mean ± SD
- 36.7 ± 17.1 cm
Accuracy and error are measured only on legal first landings on the opponent half.
Measurement details
Legal landing rate counts all attempts. Hardware accuracy is conditional on legal first landings and does not measure overall return success. Simulation uses a 30 cm accuracy threshold and hardware uses 45 cm; these rates are not directly comparable. Landing error is mean ± sample SD.
Motor self-play
Motor self-play improves landing accuracy and legal-return rate compared with prescribed continuation from the same Stage-I checkpoint and with the same additional training budget.
- Stage I
- Prescribed continuation
- Interaction-driven self-play
Evaluation protocol
30,000 rallies per checkpoint per seed, eight seeds, one fixed opponent. Curves are equal-weight seed means ±1 SEM. Rates count every return opportunity reaching each stroke index, including unsuccessful returns.
Effect of opponent position and previous target
In these controlled tests, target selection tends to shift opposite the opponent's position or the previous target.
Change opponent position
Previous target held at center.
Change the previous target
Opponent held at center.
Sampling protocol
Each map uses 500 categorical samples on the same probability scale (0–0.3). Only opponent position or the previous target changes; all other inputs stay fixed. These tests show local sensitivity, without measuring either input's contribution to winning.