Robotic table tennis system predicts ball trajectory and adapts swing in real time

Over the past decades, roboticists have introduced various systems that can replicate specific human motions and behaviors with remarkable accuracy. Some of these robots can even compete with other robots or humans in specific sports, such as the robots showcased at the RoboCup, an international robotics event at which robots play soccer with each other.

Researchers at the Massachusetts Institute of Technology (MIT)’s Biomimetic Robotics Laboratory recently introduced a new robotic table tennis platform that can successfully and rapidly hit balls with a table tennis racket. Their platform, outlined in a paper published on the arXiv preprint server, can reproduce various table tennis hit styles and spin the ball in different directions with high precision.

“The Biomimetic Robotics Lab at MIT has always strived to create performant robotic systems by innovating on both the hardware and control as seen with the Mini-Cheetah,” Kendrick Cancio, co-author of the paper, told Tech Xplore.

“We had an opportunity to create this table tennis system on behalf of the Robotics and AI Institute as a platform to explore dynamic manipulation with the ultimate goal of reaching human parity at table-tennis on a dynamic anthropomorphic platform.”

The Biomimetic Robotics Lab conducts research focusing on two primary areas of robot control: dynamic-legged locomotion (i.e., the flexible movement of legged robots) and the rapid manipulation of objects. These two areas of robotics research come with their own unique challenges.

When it comes to legged robot locomotion, a key challenge is effectively handling disturbances in the environment, while successful object manipulation entails accurately performing desired maneuvers.

“Table tennis blurs the lines between these control issues with adaptiveness and precision required as more information about an incoming ball becomes available,” David Nguyen, co-author of the paper, told Tech Xplore. “This makes it a very unique control problem that we think we can hit out of the park with our custom robotics hardware.”

The new platform developed by Nguyen, Cancio and Sangbae Kim consists of a robotic arm and a control algorithm. The algorithm can predict the path of an incoming ball and plan the actions that the arm should perform when swinging the racket to hit the ball, while also meeting specified strike conditions.

“Even while swinging, the path the arm takes is dynamically updated to ensure the paddle reaches the ball at the right location, speed, and orientation,” explained Nguyen.

“We found that planning for the entire swing rather than just the future actions was more reliable but requires more aggressive maneuvers from our arm. This is a unique advantage of our work since we can push the limits of our custom hardware far more than an off-the-shelf system.”

The robotic systems developed by the researchers has two primary components, referred to as the perception and actuation modules. The perception module consists of an off-the-shelf motion tracking system, which can localize the custom table tennis balls developed by the researchers.

“We predict the ball’s trajectory to get an anticipated strike location and strike time,” explained Cancio.

“At the same time, we have a nonlinear optimization problem which uses these values along with information on how we would like to strike the ball to generate a swing trajectory for the arm. Our model predictive controller continually solves for this arm trajectory and allows the arm to react as we get updated positions of the ball.”

The robotic arm integrated in the team’s table tennis platform is a customized version of a humanoid arm developed at MIT. This arm has high-torque and low rotor inertia, two characteristics that allow it not only to swing quickly, but also to react quickly and adapt its trajectories if initial predictions about a ball’s trajectory are wrong.

“We demonstrate an ability to adapt a trajectory to precisely meet a dynamically moving object,” said Nguyen. “Although table tennis is not going to save any lives, this kind of control could be used in difficult search and rescue situations where a more general robot like a humanoid might need to intercept an object.”

Nguyen, Cancio and Kim evaluated their robotic table tennis platform in a series of real-world experiments and found that it performed remarkably well. In these initial tests, the robotic arm hit incoming balls with a success rate of 88% and an average exit velocity of 11 m/s, following three distinct hitting styles.

“Right now, robotics is generally split between model-based approaches and reinforcement learning approaches, with some expecting the latter to be the catch-all tool in the near future,” said Cancio. “We show that constraint-based optimization still has a place for performant systems and hope to leverage the benefits of each when appropriate.”

The new system developed by this team of researchers could soon inspire other roboticists to develop similar automated table tennis platforms. In addition, Nguyen, Cancio and Kim hope to apply the hardware and control algorithm they developed to other dynamic manipulation tasks.

“Since we submitted the paper back in September, we have done a lot to improve the capabilities of the system,” added Nguyen. “Namely, we are now able to aim at specific locations on the table and plan for the entire trajectory and contact between the ball and paddle.”

As part of their future studies, the researchers plan to further enhance the table tennis robotic platform’s capabilities. By broadening the MIT humanoid arm’s workspace using a gantry (i.e., a structure supporting the arm), for instance, they could allow it to play entire games of table tennis against human users.

“We aim to keep pushing the performance of our system by expanding the workspace using a gantry and significantly increasing our ball exit velocities,” added Cancio.

“We would also like to move towards tracking standard table tennis balls to make better comparisons against humans and other robotic systems alike.”

© 2025 Science X Network