Two Unitree G1 Robots Play Table Tennis Doubles: PingpongDuo Enables Humanoid Team Coordination

PingpongDuo team submitted

Quantum Bit | Official WeChat Account QbitAI

The red-clad robot swings to hit the ball over the net, then steps aside; the blue-clad teammate steps forward to receive the next ball, plays it, and then gives way to the position.

Two Unitree G1s stand on the same side of the table, playing table tennis doubles with their human partner.

△

The two G1s on the same side of the table take turns hitting the ball, then give way to their teammate after finishing.

For robots, this 'national sport' presents a significant challenge: according to the rules of table tennis doubles, the two players on the same team must alternate hitting the ball. Even if the ball flies to one's own side, as long as it's not one's turn, the shot must be passed to the teammate.

Movement also affects the swing. As long as the body is moving sideways, the speed and direction of the racket when it hits the ball may change. Even if a single stroke is practiced to perfection, it must adapt to the body's condition after continuous movement.

Based on these situations,

researchers from Shanghai Chuangzhi College, Yuance Future, The University of Hong Kong, and NVIDIA

proposed

PingpongDuo

a method that allows two G1 robots to alternately perform striking, yielding, and returning through joint positioning planning and full-body motion control. The related paper has been published as an arXiv preprint.

△

Complete video

The positions of the two robots need to be calculated together

Each time the ball comes, the system estimates the ball's position and speed based on continuous observations, predicts its subsequent flight and bounce, then selects the position and timing for hitting the ball, and calculates the required racket speed to hit the ball to the target landing spot. This part adopts the planning logic of an existing singles system.

In doubles, it also needs to consider the teammate's position. PingpongDuo uses a team planner to simultaneously assign lateral movement targets to both robots; these targets are only passed to their respective motion controllers after risk checks.

△

Overview of the method

The planner can see a history of incoming ball positions, the current positions of both sides, and previously accepted movement targets. It uses PPO reinforcement learning to try different positioning strategies in continuous ball simulations, aiming to complete returns while learning to reduce dangerous close approaches. During the training of this layer, the lower-level motion controllers and risk model remain fixed.

The risk model is trained using another set of simulation data. The team sets random targets for the two robots and also adds disturbances to the original rule-based positioning, allowing them to execute in the simulation. During the process, if the inner side of the hands, feet, or body of both robots are less than the set distance, they are marked down to be used for training the model to identify dangerous target combinations.

The risk model estimates whether the two robots will get too close during the next movement based on the current state and candidate targets. For example, the two robots may end up squeezed together even though they are currently separated by some distance and move toward their respective targets.

If a set of targets is predicted to have a high risk, the system will adjust them. However, the two robots do not simply step back:

the robot that is preparing to strike will keep its target as stable as possible; the other robot will yield more.

The 'role weighting' imposes a higher cost on adjusting the target position and timing of the striker. The contact point and timing of this shot have already been selected. If the striker is moved away to create distance, this shot may miss entirely.

△

Team Planning: Left side illustrates goals and role switching, right side displays real-time motion.

When it's the next beat for the teammate, the priority also switches. The two robots take turns gaining striking space, while simultaneously preparing for the subsequent repositioning.

Under the same risk filter conditions, replacing the rule-based positioning with learned joint positioning increases the average number of consecutive legal returns from 3.79 to 4.87.

△

Average consecutive legal returns: rule-based positioning 3.79 times, joint positioning 4.87 times; both use the same risk filter. Error bars represent standard deviation.

Can move, can swing, but when combined, they still need to practice more.

Position is calculated, but the feet must keep up.

In real-time testing, one robot can complete

2.4

seconds of total movement distance

2.15

meters of repositioning and returning, with peak speed reaching

2.23

meters per second. This test includes stepping back, changing direction, and returning.

These stepping techniques reference human doubles actions. The research team collected full-body motion, cutting different directions and distances of movement into segments, then converting them into reference actions suitable for the robot's body structure.

When needing to move in a certain direction and for a certain distance, the system selects the corresponding segment. During training, the robot mimics the stepping techniques while learning to track target position and speed, maintain balance, and perform braking and waiting upon arrival.

Swinging is trained separately. It also controls the entire body, but the focus is on the moment of contact: the position, speed, and orientation of the racket must simultaneously satisfy the striking command. Movement requires continuous adjustment, while striking focuses precision on this brief moment; separate training allows for distinct attention to the requirements of the two tasks.

△

Short-distance and long-distance stepping, forehand and backhand swinging: action references versus real-time performance comparison.

However, switching between the two trained controllers cannot directly result in continuous doubles.

The issue lies in the moment of switching. If stepping techniques are trained from a stable posture, but in real-time it starts from a state of just finishing a swing with the body still rotating, it encounters situations not prepared for during training. Striking is similar: being able to strike from a stable position, but when switching to repositioning with the feet still moving, the action may not be suitable.

The team uses the physical state left by the previous action as the starting point for the next skill training.

Specifically, first collect the robot's posture, speed, and joint states after striking, then let the stepping controller train from these states. Conversely, also record the physical states generated during movement, allowing the striking controller to train swinging under these conditions.

In actual operation, the state machine handles alternating striking and skill switching. When the robot's striking command arrives, it can directly take over the ongoing movement; after swinging and follow-through, it returns to stepping control. The control history before the switch is retained, and the reference action undergoes a brief transition.

The two robots still use the same skill strategy. For the teammate on the other side, the system mirrors the observations, commands, reference actions, and joint actions left-right, corresponding to right-hand striking and opposite-direction repositioning.

△

First, demonstrate how the two skills switch training starting points, then observe the continuous movement and swinging in real-time.

The racket touches the ball, but it's not yet successful

In existing humanoid robots playing table tennis, racket target tracking training requires the racket to reach the target position, speed, and orientation at a specified time, without directly simulating the racket-ball interaction, nor updating the strategy based on the ball crossing the net or landing on the table. PingpongDuo incorporates racket-ball interaction simulation, using return results to train the holding arm correction.

On the existing swing motion, the system adds a strategy for the holding arm residual. 'Residual' can be understood as an additional correction amount: the original controller first provides the motion, and the new strategy makes small adjustments based on the current state of the body and the ball.

During training, the basic skills remain unchanged, and only the seven joints of the holding paddle arm learn how to be corrected.

The correction magnitude is limited and gradually added in during a period close to ball contact, then removed afterwards.

△

Holding paddle arm correction

First, the ball in the simulation should be as close to the real ball as possible. The team used motion capture equipment to record the ball's flight, table rebound, and collision with a moving paddle, and calibrated air resistance, rebound, and contact parameters in stages. Subsequent training was conducted in this simulation environment calibrated through real-world measurements.

During training, the robot first moves, then attempts to hit the ball. Contact with the ball, crossing the net, and landing on the opponent's table are all part of the training feedback. After completing a legal return, the training further examines the accuracy of the landing point and the margin over the net; missed balls and out-of-bounds shots are penalized.

The tracking of the paddle, body stability, and constraints on the correction magnitude are still retained. During training, physical parameters are also changed, and observation errors of the ball state and execution delays are added, allowing the strategy to encounter situations different from the ideal simulation.

In single-robot simulation testing, the basic skills remain unchanged. After adding the holding paddle arm correction, The legal return success rate increased from 44.2% to 55.6%, an increase of 11.4 percentage points.

The testing is arranged in a rhythm of alternating doubles, covering both moving and prepared hitting states.

△

Legal return success rates for different methods. Without refinement indicates the holding paddle arm correction is disabled, and the error bars represent the standard deviation.

The figure shows the evaluation results for five different random seeds, with 100 balls per group, and the statistics are the proportion of balls that crossed the net and landed on the opponent's table, with missed balls also counted in the total number.

The top-down view shows the results for a single group, with each method performing 100 simulated hits. The animation sequentially displays the landing and net contact positions, with a crosshair indicating the target landing point, and color representing the speed before hitting the ball. These landing results are separately counted from the average success rates of the five groups above.

△

Single group of 100 simulations: ● Legal return, ▲ Net contact, + Target landing point; color represents the speed before hitting the ball. The single group count is separately counted from the average success rates of the five groups above.

Real-world doubles coordination

Two G1 robots completed 18 consecutive alternating returns during a real-world match with a human partner.

△

Real-world continuous match: The robot team completed 18 consecutive alternating returns.

PingpongDuo is aimed at high-dynamic sports for humanoid robots, combining ball sensing, team decision-making, and full-body motion control, and achieving dual-robot coordination in real-world table tennis doubles, providing real-world verification for multi-robot collaboration in dynamic environments.

Paper title: PingpongDuo: Agile Humanoid Table Tennis Doubles with Safe Team Coordination

Project homepage: https://www.archon.tech/research/PingpongDuo

Paper link (arXiv): https://arxiv.org/abs/2610.20579

One-click triple action

"Like","Share","Heart"

Welcome to leave your thoughts in the comment section!

—

End

—

[Academic Submission] Please send an email during business hours to:

[email protected]

with the subject line [Submission], and tell us:

Who you are, Where you're from, Submission content Attach Project/homepage link, and Contact information.

🎓

We will (try to) reply to you promptly :)

🌟 Light up the star rating 🌟

Frontier advances in technology every day