MIT engineers are using powerful, lightweight design to perform robotic table tennis games to return to the camera accurately at high speed.
The new ping-pong robot includes a number of fixed robot arms that are fixed to one end of the ping-pong table and swing the standard ping-pong ball. With the help of several high-speed cameras and high-bandwidth predictive control systems, the robot quickly estimated the speed and trajectory of the incoming ball and performed one of several swing types – loops, drives or shreds – accurately hitting the ball to the desired position on the table with various rotating platforms.
During the test, the engineer threw 150 balls at the robot from the opposite side of the table tennis table. The robot successfully returned the ball with a hit rate of about 88% in all three swing types. The robot’s strike speed is close to the maximum return speed of human players and is faster than other robot ping-pong designs.
Now, the team is looking to increase the robot’s race radius so that more shots can be returned. They then envisioned that the setup could be a viable competitor in the growing field of intelligent robot training systems.
The team said that besides gaming, table tennis technology can be improved to improve the speed and responsiveness of humanoid robots, especially for searches and answers, and where the robot needs to respond quickly or expect.
“The problems we want to solve, especially those related to the really fast and precise interception of intercepting objects, can be useful when a robot has to do dynamic operations and plan its final effector will encounter the object in real time,” said David David Nugyen, a graduate student at MIT.
Nguyen is the co-author of the new study, as well as MIT graduate students Kendrick Cancio and Sangbae Kim, associate professor of mechanical engineering and head of the MIT Bionic Robot Laboratory. The researchers will present the results of these experiments in this month’s paper at the IEEE International Conference on Robotics Technology (ICRA).
Precise game
Since the 1980s, the construction robots that researchers faced were a challenge to play table tennis. This problem requires a unique combination of technologies, including high-speed machine vision, fast and agile motors and actuators, precise manipulator control, and accurate real-time predictions, and advanced planning for gaming strategies.
“If you think of various control issues in robotics, we can manipulate one end, which is usually slow and very precise, like picking up an object and making sure you master it. On the other hand, you already have dynamics and adapting to the perturbations of the system,” explains Nguyen. “Ping Pong sits between them. You’re still doing it because you have to hit the ball accurately, but you have to hit it in 300 milliseconds. So it balances with similar issues of dynamic motion and precise manipulation.”
Ping Pong robots have come a long way since the 1980s, most recently by Omron and Google DeepMind designs that use artificial intelligence to “learn” from previous Ping Pong Data to improve the performance of the robot in exchange for more and more strokes and shots. These designs have proven to be fast and precise enough to rallies with intermediate human players.
“These are professional robots designed to play ping pong,” Cácio said. “With our robots, we are exploring how the technology used using ping pong can be translated into a wider range of systems, such as humanoid or anthropomorphic robots, that can do many different, useful things.”
Game Control
For their new design, the researchers modified a lightweight, high-power robot arm, and Kim’s lab is part of the MIT humanoid animal – a bipedal, two-armed robot, about the size of a child. The team is using robots to test a variety of dynamic operations, including driving on uneven and changing terrain as well as jumping, running and performing tailgate, with the aim of one day deploying such robots for search and answering operations.
Each humanoid animal has four joints or degrees of freedom in the arms, each of which is controlled by an electric motor. Cancio, Nguyen and Kim built a similar robotic arm that they adapted to Ping Pong by adding a degree of freedom to their wrist to control the paddle.
The team secured the robot arm to one end of the standard ping pong table and set up a high-speed motion capture camera around the table to track the robot’s bounce ball. They also developed optimal control algorithms that predict the speed and paddle direction of the arm according to principles of mathematics and physics to use a specific type of swing: loop (or topspin), drive (direct drive), or CHOP (CHOP (Backspin)) to hit the incoming ball.
They implemented the algorithm using three computers that simultaneously processed camera images, estimated the real-time state of the ball, and converted these estimates into commands of the robot motor to react quickly and swing.
After bounced 150 consecutively, they found that the robot’s hit rate or return ball accuracy was the same in all three swings: 88.4% of loop strike, 89.2% of chopped, and 87.5% of drives. Since then, they have adjusted the robot’s reaction time and found faster arm hitting speeds with existing systems at 20 meters per second.
In their paper, the team reported that the robot’s strike speed or paddle hitting speed averaged 11 meters per second. As we all know, senior human players will return to the ball at a speed of 21 to 25 meters. Since writing the results of the initial experiment, the researchers further adjusted the system and recorded strike speeds of up to 19 meters per second (about 42 miles per hour).
“Some of the goals of the program are to say we can achieve the same athletic ability as people,” Nguyen said. “We’re really close in terms of strike speed.”
Their follow-up work also aimed at the robot. The team incorporated control algorithms into the system, not only to predict how to hit the ball. With its latest iteration, researchers can set target positions on the table and the robot will hit the ball to the same position.
Because it is already fixed to the table, the robot’s mobility and access is limited, and most of it can be returned to the ball in the crescent-shaped area around the midline of the table. In the future, engineers plan to perform robots on gantry or wheeled platforms, allowing them to cover more tables and return to a wider range of lenses.
“Given the way your opponent hits, one of the big things about ping-pong is predicting the ball’s rotation and trajectory, which is the information that the automatic ball launcher won’t give you,” Canceo said. “A robot like this can mimic the movements the opponent will do in the gaming environment, thus helping humans play and improve.”
This study was partially supported by the Institute of Robotics and AI.
