Time lapse photos show a new ping-pong-playing robot performing a top spin. The robot quickly estimates the speed and trajectory of an incoming ball and precisely hits it to a desired location on the table. (Credit: Courtesy of David Nguyen, Kendrick Cancio and Sangbae Kim)
In a nutshell
- MIT’s custom robot arm can perform human-style table tennis shots, including topspin loops, flat drives, and backspin chops, with an impressive 88% success rate across styles.
- The robot’s paddle accelerates up to 300 m/s² and can return balls at speeds up to 14 meters per second (31 mph), approaching the level of intermediate human players.
- A fast-reacting predictive system lets the robot strike with precision in real time, using high-speed cameras and model predictive control to respond within 7.5 to 16 milliseconds of seeing the ball.
CAMBRIDGE, Mass. — Scientists at the Massachusetts Institute of Technology have created a robot that can play table tennis with the skill and flair of a human player. The robot smashes returns at speeds of up to 14 meters per second (about 31 mph), fast enough to challenge intermediate human players.
This custom robotic arm is specifically designed for the quick, precise movements needed in table tennis. Unlike previous robotic attempts at the sport, which often used off-the-shelf industrial robots that were too heavy and slow, this purpose-built machine achieves impressive speed using lightweight components with motors positioned to minimize inertia.
The research team says their creation is capable of executing various hit styles with impressive precision, power, and consistency—all crucial elements for table tennis play. Their work is scheduled to appear at the IEEE International Conference on Robotics and Automation (ICRA) 2025.
And it doesn’t just have impressive speed. The system can perform different types of shots that human players use – loops with topspin, flat drives, and chops with backspin – with an 88% success rate across all three styles. This versatility moves robotic table tennis closer to the fluidity and adaptability of human play.
The robot’s custom arm weighs just 3 kilograms (6.6 pounds) but packs four high-torque motors that can accelerate the paddle at 180-300 meters per second squared, far beyond what standard industrial robots can achieve. This allows it to react quickly enough to meet incoming balls with the correct paddle position, orientation, and velocity.
According to the researchers, top human players can hit forehand loops at about 21 meters per second (47 mph) and drives at around 25 meters per second (56 mph). While the MIT robot hasn’t quite matched those speeds yet, it’s getting closer than previous attempts.
The system relies on a sophisticated prediction algorithm that calculates where an incoming ball will be and when it will arrive at the striking plane. Six motion capture cameras track reflective balls at 120 frames per second, while specialized software predicts the ball’s trajectory, accounting for bounce dynamics and air resistance.
To generate the robot’s swinging motion, the team developed a model predictive control (MPC) system that constantly replans the trajectory based on updated ball position information. This allows the robot to adjust mid-swing if the prediction changes.
The researchers tested their robot with 150 balls for each type of shot – loop, drive, and chop. The success rates were consistent at 88.4% for loops, 89.2% for chops, and 87.5% for drives. Average ball exit velocities measured around 11 meters per second (25 mph).
The researchers note that table tennis presents unique challenges compared to other tasks, requiring intentional, impulsive contact executed with exceptional speed and accuracy, all while simultaneously controlling the robot and predicting the ball’s path.
Previous table tennis robots have struggled. Some required large systems that moved the paddle above the table, but couldn’t be easily adapted for other tasks. Others used industrial robotic arms that were too heavy to accelerate quickly enough for fast-paced play.
The team’s custom five-degree-of-freedom arm resembles a human arm in structure, with shoulder and elbow joints. Four powerful U10 actuators provide 34 Newton-meters of torque, while a fifth smaller motor controls the paddle orientation. By placing the heavy motors closer to the base and using lightweight materials elsewhere, they minimized the arm’s inertia to enable rapid acceleration.
For the robot to succeed, timing is everything. The system must receive camera data, predict the ball’s path, optimize a striking trajectory, and execute the motion – all within milliseconds. The researchers measured their system’s reaction time between receiving new ball observations and executing new trajectories at just 7.5-16 milliseconds.
The paddle’s position, orientation, and velocity must all be precisely controlled to achieve different types of shots. For topspin loops, the paddle approaches from below with a 45° upward angle. For chops, it strikes from above with an 18° downward angle. Drives use a neutral flat paddle position.
While impressive, the system could still be improved. The researchers note that their current setup restricts incoming ball speeds and doesn’t yet incorporate spin detection for incoming balls. The robot also currently focuses only on striking the ball, not on the strategic placement of returns.
Future improvements will include automated shot aiming, spin estimation, and expanding the robot’s workspace to cover the entire table. These enhancements would bring the system closer to playing full competitive matches against human opponents.
MIT’s table tennis robot isn’t just built to pick things up; it’s designed to move fast and react in real time, just like in a real game. This shift promises more capable robots for sports, manufacturing, and potentially everyday environments where things rarely stand still.
Paper Summary
Methodology
The researchers developed a custom 5-degree-of-freedom robotic arm specifically designed for table tennis, weighing only 3kg with four U10 actuators that provide 34 Nm of peak torque. To track balls, they used six OptiTrack Flex 13 motion capture cameras at 120Hz that tracked retro-reflective tape-wrapped table tennis balls. The system uses three computers communicating via Lightweight Communications and Marshalling (LCM) to minimize reaction time. For ball trajectory prediction, they implemented a simplified dynamics model that integrates air resistance and bounce mechanics. The robot’s motion control uses an optimal control problem (OCP) formulated within a fixed-horizon model predictive controller (MPC) that continuously replans trajectories based on updated ball positions.
Results
The robotic system achieved an 88% success rate across three different shot types: loops (topspin), drives (flat), and chops (backspin). The average exit velocity of the balls was 11 m/s, with peak velocities of up to 14 m/s during testing. Position errors were generally within the 7.5 cm critical distance (paddle radius), and velocity magnitude was typically within 2 m/s of the desired strike speed. The system demonstrated better control over vertical angles than horizontal ones due to the wrist joint configuration. The effective reaction time between receiving new ball observations and executing new trajectories was measured at 7.5-16 ms.
Limitations
The researchers identified several limitations in their current implementation. First, their Fixed Horizon MPC approach means the current arm state isn’t guaranteed to be close to the new calculated solution, potentially resulting in large trajectory adjustments. Second, the arm needs additional control authority at the wrist to more precisely control certain strike angles. Third, the system is currently constrained to striking in a single plane rather than full 3D positioning. Finally, the prediction system’s variable computational requirements created latency issues that restricted the incoming ball speeds that could be tested successfully.
Funding/Disclosures
The research was supported by The AI Institute in Cambridge, Massachusetts. All authors are affiliated with the Biomimetic Robotics Laboratory at Massachusetts Institute of Technology (MIT).
Publication Info
The paper “High Speed Robotic Table Tennis Swinging Using Lightweight Hardware with Model Predictive Control” by David Nguyen, Kendrick D. Cancio, and Sangbae Kim has been submitted to appear in IEEE ICRA 2025 and was published on arXiv on May 2, 2025.
