Table Tennis Robot: From Ball Machines to ABB Arms

A table tennis robot has to see a ball, predict its spin and hit it back within a fraction of a second, so the game is a hard test for robots and a useful one for training people. This post follows sports training machines from the ball launchers of the 1920s to the first robot ping-pong players and Omron's FORPHEUS coach. It then covers the research robots that Google and DeepMind built on ABB arms, the humanoids that played table tennis at the World Humanoid Robot Games in Beijing this August, and what home and club training robots could look like next.

A visitor plays against Omron's table tennis robot at Hannover Messe 2016. Photo: Bahnfrend / Wikimedia Commons (CC BY-SA 4.0)

From ball launchers to the first robot ping-pong players

Sports training machines started as simple launchers. In 1897 the mathematician Charles Howard Hinton designed a gunpowder-powered pitching machine for batting practice at Princeton University. It could vary its speed and threw curve balls using two rubber-coated steel fingers at the muzzle. In tennis, René Lacoste built a hand-cranked ball machine he called the lance-balle so he could practise alone, and in March 1928 press photographers watched it being tested at the tennis club in Asnières, near Paris. Machines like these fed balls at a set rhythm and could not react to anything the player did.

A ball-throwing machine tested at the Asnières tennis club by René Lacoste, March 1928. Photo: Agence Rol / Bibliothèque nationale de France via Wikimedia Commons (public domain)

Robots that could return a ball came out of research labs in the 1980s. In 1983 Billingsley proposed a robot table tennis competition in the magazine Practical Computing and set out simplified rules for it. At AT&T Bell Laboratories, Russell Andersson built what later researchers describe as the first robot ping-pong player able to play against humans and machines, presented in 1988. It used a high-speed video system and a six-axis PUMA 260 industrial arm holding a paddle on a 0.45 m stick. Google and DeepMind would use the same combination of fast cameras and a stock industrial arm more than 30 years later.

A Unimate PUMA 500 robot from 1983 at the Deutsches Museum; Bell Labs' ping-pong robot used a smaller PUMA 260. Photo: Theoprakt / Wikimedia Commons (CC BY-SA 3.0)

Robot opponents then moved from the lab to trade fairs. The Vietnamese company TOSY built TOPIO, a humanoid robot made to play table tennis, and showed version 3.0 at the International Robot Exhibition in Tokyo in 2009. Omron built its first table tennis robot in 2013 and unveiled it at CEATEC in Japan the next year. Later versions of FORPHEUS tracked the ball with cameras 80 times a second, projected onto the table where the ball would land, and rated each player's skill to pitch the rally at the right level. In 2016 Guinness World Records certified FORPHEUS as the first robot table tennis tutor.

Omron's FORPHEUS table tennis robot rallying with a visitor at Hannover Messe 2016. Photo: Bahnfrend / Wikimedia Commons (CC BY-SA 4.0)

How Google, DeepMind and Sony built robots that play to win

The next step used ordinary industrial arms. In 2022 Google researchers published GoalsEye, a system that learned to hit table tennis balls to chosen targets. Its player robot was an ABB IRB 120T six-axis arm on a two-axis Festo linear actuator, and the arm took joint and velocity targets about 240 times a second through ABB's Externally Guided Motion (EGM) interface. In August 2024 Google DeepMind went further with a robot built on an ABB IRB 1100 riding two Festo gantries, 4 m across the table and 2 m towards it. It won 13 of 29 matches against people: every match against beginners, 55% against intermediate players and none against advanced ones.

Students program an ABB IRB 120, the robot family Google used for its GoalsEye table tennis player. Photo: Peterantala / Wikimedia Commons (CC BY-SA 3.0)

Sony AI's Ace, described in Nature on 22 April 2026, is a custom robot with eight joints, nine cameras that fix the ball's 3D position and event-based vision sensors that measure spin. It won three of five matches against elite amateur players, and in later matches in December 2025 and March 2026 it beat professionals. In Beijing, the second World Humanoid Robot Games added table tennis this August. Twelve teams played on the same full-size humanoid with no remote control, and the PKU-BAAI team beat the University of Hong Kong's SMASH team 2-0 in the final. Tang Minqin, an organiser of the table tennis event, told the Global Times the robots may have only about 0.3 seconds to perceive the ball, predict its path, decide and swing.

TOPIO 3.0, a humanoid table tennis robot by Vietnam's TOSY, at the International Robot Exhibition in Tokyo, 2009. Photo: Humanrobo / Wikimedia Commons (CC BY-SA 3.0)

ABB does not sell a table tennis robot, but its small arms keep turning up in this research. The IRB 1100 carries 4 kg with a reach of 475 or 580 mm. ABB introduced it in 2018 as its smallest and lightest robot yet, with up to 35% faster cycle times than before and repeatability of 0.01 mm. The larger IRB 1300 carries 7 to 12 kg with up to 1.4 m of reach and up to 27% faster cycle times than the generation before it. EGM lets outside software steer the arm in real time. DeepMind rehearsed its rallies in the MuJoCo physics simulator, and ABB users test motion the same way in RobotStudio before the real arm moves.

An ABB IRB 1300 Clean Room robot. Photo: ABB

Where sports training robots go next, at home and in clubs

Training robots for amateurs are changing quickly. At IFA in Berlin in September, ACEMATE presented the S10 Pro tennis robot, which tracks the ball with dual 4K cameras, moves around the court, feeds balls at up to 120 km/h and has a Battle mode in which it plays points as an AI opponent. It costs $2,999. On 23 September TENNIIX and LimX Dynamics showed ULTRA MAX, a tennis robot on a wheeled-leg platform, at the Billie Jean King Cup Finals, in a demonstration with the Chinese player Wang Qiang.

A boy loads a wheel-type pitching machine at a youth baseball field. Photo: jgodfrey / Wikimedia Commons (CC BY 2.0)

Safety and cost are the main open problems. An arm that swings a paddle fast enough to beat an elite player can hurt anyone who steps into its path, so it needs a guarded zone in the same way industrial arms do in factories. Today's home products launch balls instead of swinging at them. A $2,999 tennis robot is still a purchase for clubs and keen players, and an outdoor court adds dust and rain. ABB's IRB 1300 comes in a Foundry Plus 2 version sealed against water and dust, which is what an outdoor training rig would need, though no such ABB product has been announced.

An ABB IRB 1300 with Foundry Plus 2 protection, built to handle water and dust. Photo: ABB

The research also changes what people building sports robots need to know. DeepMind trained its player in simulation and then moved the skill to the real arm, and Sony used model-free reinforcement learning, so students in this field need vision, control and machine learning together, plus time on real hardware. ABB pitches its smallest industrial robot, the IRB 1010, at schools and labs for exactly that kind of practice. Factories gain from the same work, because seeing a moving object and reacting in time is also what a robot needs for picking parts from a moving line, the job covered in conveyor tracking.

A class works with ABB's IRB 1010, which ABB pitches for education. Photo: ABB

Innovation

AI solutions for effortless ABB robot control.

Automation

Robotics

ceojohntran@chatwithrobot.net

+84905311611

© 2025. All rights reserved.

qtran1215@gmail.com