The History and Evolution of Multiball Training in Table Tennis

From Japanese volleyball training and washbasins full of balls in 1960s China to programmable robots, ball-recycling systems, electronic targets and app-driven performance analysis, multiball training has evolved by removing one training limitation after another.

Today, multiball is such a familiar part of serious table tennis training that it is easy to forget how revolutionary the idea once was.

Practising coach-fed multiball with Fastpong interactive targets

A coach stands beside the table with a box or basket of balls and feeds them continuously to a player. Instead of waiting for a rally to develop naturally, the coach can dictate the speed, spin, placement and frequency of every shot.

A player can practise the same movement hundreds of times, rehearse a particular combination, work on movement between two positions, or be pushed physically far harder than would normally be possible in a rally. Then came the table tennis robot.

Modern table tennis robot

Robots removed the need for another person to feed the balls. Later machines added different spins, oscillation, programmable sequences, ball recycling, digital controls and smartphone apps. But throughout most of that evolution, one part of the training process remained largely untouched: The robot knew what ball it had sent, but knew almost nothing about what the player sent back.

That distinction helps explain both the history of multiball and where table tennis training technology is heading next.

Where did table tennis multiball come from?

One of the best-documented accounts of the origins of modern table tennis multiball comes from China — but it begins with a Japanese volleyball coach. Hirofumi Daimatsu (大松 博文, Daimatsu Hirofumi) was the legendary coach of the Japanese women’s volleyball team that became known internationally as the “Oriental Witches”. His teams were famous for an exceptionally demanding training regime based on enormous numbers of repeated actions. He subsequently coached Japan to the women’s volleyball gold medal at the 1964 Tokyo Olympics.

Chinese table tennis coach Liang Youneng (梁友能), who later became an ITTF Merit Award recipient,was among those who studied Daimatsu’s methods. In a CCTV history of Chinese table tennis, Liang recalled that Chinese Premier Zhou Enlai (周恩来) encouraged the national team to learn from the training style of Daimatsu’s Japanese women’s volleyball team.

Liang adapted the intensive multiball concept to elite table tennis training. The coaches obtained washbasins, filled them with table tennis balls and began imitating the multiball training method. Chinese world champion Zheng Minzhi (郑敏之) remembered Liang feeding basket after basket of balls, changing the placement left and right, long and short, while the players attempted hundreds — sometimes around a thousand — repetitions.

The Chinese national team’s adoption of this method during the 1960s was enormously important to its development and subsequent spread through international table tennis.

Why multiball changed practice

Traditional rally practice contains an unavoidable limitation: the next ball depends on the previous ball being returned successfully.

If a player wants to practise 100 forehand loops, a conventional rally requires someone to create 100 suitable balls — and every mistake interrupts the sequence. Multiball separates training repetition from rally continuity.

The feeder can send:

• exactly the same ball repeatedly;
• balls to alternating locations;
• different spins in a predetermined pattern;
• increasingly difficult combinations;
• completely random placements;
• or balls at a frequency designed primarily for physical conditioning.

A good coach can also watch the player and modify the next feed instantly. Human multiball therefore remains extraordinarily valuable even in the age of sophisticated robots, but the coach has to be present, and having a dedicated full-time coach is not something available to most players. Good multiball feeding is itself a learned skill. Feeding hundreds or thousands of balls is tiring, and a coach concentrating on the feed cannot always observe the player as closely as they would like.

The obvious question was: could the feeding be automated?

Before the modern robot: the vacuum-cleaner solution

In 1946, American inventor George Lemon filed a patent for an “Apparatus for Practicing Ping-Pong and the Like”. The patent was granted in 1950.

One of George Lemon’s 1950 patent drawings

It is one of the more entertaining branches in the family tree of the modern table tennis robot. Lemon proposed a large catching screen, a ball collection system and a tube that would propel balls back towards the player using compressed air, suggesting the use of an ordinary tank-type household vacuum cleaner.

This 1950 design already contained two ideas that would reappear decades later:

1. a “robot partner” for a player practising alone;
2. and collecting returned balls so they could be used again.

The problem was ball control. Pneumatic systems can move a lightweight table tennis ball very effectively, but accurately controlling velocity, trajectory and particularly spin is much harder.

Later patent literature describing Lemon’s design noted problems associated with reliable ball feeding and maintaining air pressure. The eventual answer was not more sophisticated air pressure. It was rubber wheels.

Joseph Newgarden and the mechanical table tennis robot

One of the most important names in the history of the commercial table tennis robot is Joseph E. Newgarden Jr. Newgarden began experimenting with a table tennis robot as a hobby in 1972. In April 1973 he filed a patent for a “Ball Projecting Device with Spin Producing Mechanism”, published in February 1974.

One of Joseph Newgarden’s table tennis robot patent drawings

The patent contains much of the mechanical DNA that remains visible in table tennis robots today:

• a hopper containing many balls;
• a mechanism feeding balls individually;
• rotating wheels gripping the ball;
• control of ball frequency;
• control of speed;
• and manipulation of the throwing mechanism to alter spin.

Rather than trying to steer a ball using air, the machine made controlled physical contact with it — an engineering decision that has stood the test of time. But transforming a prototype into a dependable consumer product was not easy. Newgy says Newgarden spent approximately 16 years on research and development before the first commercial Robo-Pong appeared in 1988.

1988: Robo-Pong becomes a commercial product

The Newgy Model 1929 entered production in 1988 and remained in production until 1994. It was followed by the Robo-Pong 2000 in 1994, the Robo-Pong 1000 later in the decade, and subsequently the 1040, 2040 and digital generations of Robo-Pong.

Newgy Model 1929

The ball collection net was also improved in subsequent generations, although not all table tennis robots adopted continuous ball recycling.

Robo-Pong advertisement, Table Tennis Today, September/October 1994

Hopper robots versus ball-recycling robots

At the inexpensive and portable end of the robot market, a common design is simply: hopper → feeder → throwing head

Typically 50–100 balls are loaded into the hopper depending on its capacity. Once they have been fired, practice stops while the player walks around the table, collects the balls and reloads the machine. The obvious attraction is price and portability. Removing the catching net, return channels and recycling mechanism makes a robot smaller, simpler and substantially cheaper.

For occasional practice this may be an acceptable compromise. For serious multiball training, however, the limitation quickly becomes apparent. A robot firing 50 balls per minute will empty a typical hopper in a minute or two. A high-volume training session can easily involve thousands of balls. Without recycling, that potentially means constantly interrupting training to collect and reload balls, which diminishes one of the principal advantages of multiball: sustained repetition. The low purchase price makes them attractive as an entry point, but the inconvenience of repeatedly collecting and reloading balls can make them much less appealing for regular, high-volume training.

Simple hopper feed

At the higher end, manufacturers increasingly added a catching and recycling net. Returned balls hit the net, fall into a collection channel and are fed back into the robot. The same set of balls can therefore circulate through exercise after exercise.

Recycling systems are larger, more complicated and more expensive, but for sustained robot training, automatic recycling is not merely a convenience; it fundamentally changes how practical the robot is to use.

1994: the robot starts watching the return

In 1994, Newgy was advertising Pong-Master, describing it as an electronic interactive table tennis game. Instead of merely returning balls from the robot, the player had to hit physical sensor targets placed on the table. The system used three different-sized circular sensor targets that could be positioned around the table. It detected successful target hits and kept score.

The electronics available in a 1990s consumer product were relatively limited. Pong-Master was essentially a scoreboard connected by wires to three discrete targets, rather than a complete training-data system. Conceptually, however, it was far ahead of its time. It was one of the first attempts to make robot practice interactive rather than simply automated.

The next problem: one ball is not a rally

A conventional robot can be extremely good at repetition, but repetition can also become its greatest weakness. Consider a four-ball exercise:

forehand topspin → forehand push → backhand topspin → backhand push

Each ball requires its own speed, spin, trajectory and placement. While a human multiball coach can provide such a sequence, the traditional robots could not. Some machines oscillated from left to right but that was a far cry from exact individual placements, arcs and spins. This deficit drove the next major stage in robot evolution.

TTmatic, Amicus and programmable ball sequences

By the beginning of the 2000s, increasingly sophisticated robots were beginning to reproduce not merely repeated strokes but complete training patterns. A 2001 advertisement for the German TTmatic 402 and 402B, for example, described electronic programming and a rotating head which could be turned between chop, sidespin and topspin, combined with a large automatic ball-return system.

Butterfly’s Amicus 3000, marketed by the early 2000s, took programming considerably further. A 2003 advertisement challenged players with the question: “CAN YOUR ROBOT DO THIS?” It demonstrated a five-ball sequence containing a short no-spin serve, heavy backspin, light underspin, heavy topspin and a fast attacking ball. Individual balls could have different speed, spin and placement settings, and programmes could be saved in memory.

Amicus 3000 advertisement, USA Table Tennis, March/April 2003

The early Amicus 3000 used a two-wheel system. The later Amicus family eventually moved to the three-wheel throwing head that is associated with the brand today. That distinction leads to an interesting ongoing engineering debate between table tennis robot manufacturers.

One wheel, two wheels, three wheels — or two complete heads?

Robot manufacturers have taken several different approaches to controlling spin.

One-wheel systems

Some simpler robots use one powered throwing wheel and a friction surface or second passive contact point. These can be mechanically straightforward and remarkably durable.

Changing from topspin to backspin or sidespin, however, generally requires changing the orientation of the throwing head. They are excellent at repeatedly producing one configured ball, but much less suited to instantly changing between radically different spins in a programmed rally.

Two-wheel systems

Two independently powered wheels give the designer much greater control. If both wheels run at similar speeds, the robot can create a comparatively low-spin ball. Changing the relative wheel speeds creates topspin or backspin. To generate sidespin, the axis of the wheels can be rotated.

Older robots required this to be done manually. More sophisticated modern two-wheel machines can motorise the rotation of the head, allowing the robot itself to change the spin axis during an exercise. This architecture therefore remains very much alive.

OUKEI two-wheel throwing head

It competes directly with the three-wheel approach rather than having been superseded by it.

Three-wheel systems

The alternative is to place three independently driven wheels around the ball. Modern Butterfly Amicus and Power Pong machines are prominent examples. The great theoretical advantage is that almost any spin axis can be created by varying the three wheel speeds without rotating the complete throwing head. Butterfly explicitly promotes this as an advantage: the robot can move between topspin, backspin and sidespin combinations quickly while a lightweight deflector controls placement.

Butterfly three-wheel throwing head

But there is an engineering trade-off. Three wheels mean that the ball’s release depends on the geometry and condition of three simultaneous contact surfaces.

Butterfly’s own Amicus maintenance instructions acknowledge that as the wheels wear, the clearance between them changes. The result can be balls being thrown to irregular lengths, requiring the wheel spacing to be adjusted. In practice, users have also reported repeatability and maintenance problems, including individual balls landing noticeably away from the intended position. Shot-to-shot consistency therefore remains a legitimate concern with three-wheel designs, particularly as the wheels wear and their spacing changes.

Rotating two-wheel heads also have their own issues with motors, bearings and a comparatively large moving mass of their own. The two architectures represent different engineering compromises.

Three wheels: change spin electronically without rotating the throwing assembly. Two wheels + rotating head: use a simpler wheel geometry but mechanically rotate the spin axis. Neither approach has eliminated the other.

The wonderfully excessive solution: use two heads

There is another way to solve the problem. If a single throwing head cannot change configuration fast enough between two very different balls… use two throwing heads. This has produced some of the strangest-looking — and technically interesting — robots in table tennis. Manufacturers include OUKEI (奥奇), officially Zhongshan Oukei Sporting Supplies Co., Ltd., (中山市奥奇体育用品有限公司), Y&T and Double Dragon.

Typically, each head can be configured differently, allowing the machine to jump rapidly between different types of ball without asking one throwing assembly to mechanically transform itself between shots.

STIGA-branded dual-head robot

They may look slightly like a piece of laboratory equipment waiting to perform an experiment on the player, but the logic is perfectly sound. The trade-off is substantially greater mechanical complexity: two throwing heads, additional motors, multiple ball-feed paths and more components requiring adjustment. In practice, that additional complexity also creates more potential points of failure.

China’s expanding robot industry

By the 2000s and 2010s, China had become a major source of table tennis robots at almost every price point. Alongside OUKEI were manufacturers such as Y&T, whose robots have appeared internationally both under the Y&T name and through distributors and rebranding arrangements. Models ranged from very inexpensive hopper-fed machines to large programmable robots with catching nets and ball recycling.

Table tennis robots were no longer restricted to national training centres or wealthy clubs. Machines gradually became affordable for ordinary club players and even home users. At the same time, the high end continued to become more sophisticated. Butterfly, Power Pong, Newgy, OUKEI, Y&T and other manufacturers were now competing in an increasingly sophisticated market.

The smartphone robot boom

The arrival of ubiquitous smartphones created another obvious evolutionary step. Why not program the exercise visually on a phone rather than requiring a box covered with buttons, switches and numerical displays? Numerous manufacturers and startups pursued that idea.

Trainerbot

In 2016, Trainerbot appeared on Kickstarter promising a small, smartphone-controlled table tennis robot. Its crowdfunding campaign attracted 683 backers and more than $250,000, demonstrating the demand for a compact programmable machine.

Part of its philosophy was portability. Its designers deliberately avoided a large recycling system and instead used a small hopper.

The project also demonstrated the other side of hardware crowdfunding: building an impressive prototype is considerably easier than manufacturing thousands of reliable machines. Delays stretched far beyond the original timetable, and Trainerbot never became the mainstream robot platform that its successful campaign initially suggested it might become.

Trainerbot in action

JOOLA Infinity

JOOLA Infinity followed on Kickstarter in 2019. Its concept combined an app-controlled two-wheel robot with programmable drills and a catching/recycling net.

Again, the robot was gradually becoming software-defined. But Infinity also remained a relatively short-lived branch of the robot family tree rather than establishing the long-term market presence of Robo-Pong or Amicus.

JOOLA Infinity

PongFox: another software-controlled robot

PongFox, based in Bengaluru, India, also took up the challenge in 2018 of producing a software-defined robot. The founder Kiran Kumar started with a two-head Arduino prototype and eventually settled on a three-wheel implementation, which he began shipping in 2020.

PongFox, note the large hopper

A completely different approach: the robot that actually uses a bat

Almost every table tennis training robot ultimately solves the same engineering problem in roughly the same way: one or more rapidly rotating wheels grip the ball and propel it towards the player.

From a purely mechanical point of view, spinning wheels are a much simpler way of launching table tennis balls, but they do not look or behave exactly like a player or coach. A human feeder creates spin through the direction, speed and angle of the racket at impact. There is also visible movement before contact that gives the receiving player information about the ball that is coming.

The PONGBOT M-ONE (庞伯特 M-ONE) tries to reproduce human multiball much more literally. Instead of firing the ball through spinning wheels, M-ONE has a humanoid-style pair of mechanical arms. One arm releases the ball and the other strikes it with an actual table tennis racket, attempting to imitate the feeding action of a human coach.

The robot can vary speed, spin and landing position and can be controlled through a mobile app. Its developers also combined it with trajectory analysis and motion-capture systems capable of providing information about ball speed, height over the net and player movement.

The project grew out of work at the China Table Tennis College at Shanghai University of Sport, together with Chinese robotics company SIASUN Robot & Automation. The college established an artificial-intelligence research centre in 2015, and an earlier robot produced by the collaboration appeared at the China International Industry Expo in 2017. After another two years of development, M-ONE was officially launched in China in March 2020.

PONGBOT M-ONE robot striking the ball with a racket

Moving a complete robotic arm and racket for every ball is mechanically much more demanding than spinning a pair of small wheels. The physical machine is much larger and more expensive than a traditional table tennis robot and therefore unlikely to replace compact wheel-based machines for ordinary home practice.

The earliest Chinese multiball coaches stood beside the table with containers full of balls and physically struck each ball towards the player. M-ONE represents an alternative evolutionary path in which the robot tries to become the human multiball coach, rather than simply becoming an increasingly sophisticated ball launcher.

Increasingly capable robots

Today's table tennis robotsrange from simple hopper-fed machines to sophisticated programmable systems. By the 2020s, a high-end table tennis robot could do things that would have seemed extraordinary to George Lemon or Joseph Newgarden:

• choose individual placements;
• independently control speed and spin;
• produce serves and rally balls;
• switch between topspin and backspin;
• introduce randomness;
• store complex exercises;
• run them from a phone or tablet;
• and recycle balls almost indefinitely.

But there was still one major piece missing

Even after all these improvements, most table tennis robots remained fundamentally open-loop machines. They could precisely control the ball being sent — for example delivering heavy backspin to the forehand — but had little or no information about the player’s response. Such systems could not tell whether the player had successfully pushed it to the opponent’s backhand corner, returned it to the middle of the table, or missed completely.

That was an important drawback of robot practice. When playing against a real opponent, every shot should have an intended destination. During robot training, however, it is very easy to become completely focused on the incoming ball and neglect the return target.

The robot had automated the feeder, but not the complete training loop.

From ball delivery to interactive training

Newgy’s Pong-Master recognised this problem surprisingly early. Closing that loop requires a different model:

stimulus → action → measured response → feedback

rather than:

ball → player → next ball

Fastpong took this closed-loop approach considerably further.

Fastpong: treating the robot as part of a system

Fastpong’s initial commercial product, released in 2023, consisted of interactive electronic tiles that simply sat on the table surface and could easily be removed when not in use. A sensor attached to the player’s existing robot detected each ball as it was launched, allowing the targets to synchronise with conventional robots from different manufacturers. The tiles could be programmed as return targets, providing immediate visual feedback while the system collected data about the player’s performance.

That changed the architecture of robot practice. A conventional robot largely operates in one direction.

Programme → Robot → Ball → Player

Fastpong extends the chain:

Programme → Robot → Incoming Ball → Player → Return Target → Detection → Feedback → Performance Data

As the robot sends the ball, a tile can illuminate to tell the player where the return should go. That means the player must process two pieces of information:

1. What ball am I receiving?
2. Where am I supposed to play it?

The return is then detected, allowing the system to record whether the target was hit rather than merely knowing that another ball was launched.

But increased capability created another problem: setup complexity.

A complex drill might require several minutes of entering wheel speeds, arc angles, placements and timings. Some robots cannot produce complex sequences at all. Others can, but configuring them can become an exercise in programming the robot rather than practising table tennis.

The first Fastpong system therefore solved the return-feedback problem, but it exposed another limitation: the robot and the targets still had to be configured separately. That problem was compounded by the increasing complexity of the more capable robots.

The next step: full integration

Fastpong recognised this problem and, in 2026, released a fully integrated system designed around the needs of players and coaches. The substantially expanded mobile application now controls both the robot and the interactive tiles, while also bringing together built-in lessons, games, exercises, statistics, performance analytics and Smart Class Management.

Designed around ease of use for players and coaches, the new system shifted the emphasis from what the hardware could theoretically do to how easily the player could tell it what to do.

Return to the four-ball exercise

forehand topspin → forehand push → backhand topspin → backhand push

Each of those four balls can require:

• a different incoming spin;
• a different incoming speed;
• a different incoming placement;
• a different stroke from the player;
• and a different intended return placement.

This is much closer to what a coach means by an exercise than simply telling a robot to oscillate between two positions. In Fastpong, the mobile app configures the integrated robot and corresponding tile targets as part of the same exercise.

The player therefore no longer has to translate the exercise into a collection of robot motor settings. They simply select the training they want to do, and the system configures the equipment. For practical examples of how these systems can be used in training, see our table tennis robot drills for better match play.

Measuring the result

The integration also makes training data more meaningful. Rather than recording only how many balls were delivered, Fastpong builds statistics around what the player actually did.

The system can provide real-time landing feedback and record accuracy and return speed, and show where missed balls land, including heatmap-based analysis. Sessions can then contribute to longer-term performance statistics rather than disappearing as soon as the system is switched off.

The key difference is that ball delivery, visual instruction, return detection and performance analysis are integrated into the same training system.

The evolution of multiball at a glance

More than sixty years of removing limitations

What comes after the table tennis robot?

The original multiball pioneers were not trying to invent a robot. They were trying to create better practice. The washbasin full of balls was simply the technology available to them. The development path has been long and winding. There have been crowdfunded smartphone robots that looked revolutionary but struggled to make the transition from prototype to mass production. There have also been ideas that arrived decades before the rest of the technology was ready for them. The first robots automated repetition. Ball-recycling systems removed the interruption of collecting balls. Programmable robots automated variation. Two-wheel rotating heads, three-wheel systems and dual-head machines allowed increasingly complex combinations. Smartphones made programming easier.

The next logical stage is for the training equipment to understand not only what it asks the player to do, but also what the player actually does. That means giving the player a return target. Detecting the result. Measuring accuracy and speed. Showing where mistakes went. Recording improvement over time. And configuring the whole exercise without forcing the player to think like a robot engineer. That is the direction in which Fastpong has evolved.

Thinking of Fastpong simply as another table tennis robot misses the full picture. The robot is now one component of a wider table tennis training system. Multiball began by allowing a coach to send another ball immediately. More than sixty years later, the next stage of its evolution is increasingly about understanding what happens after that ball comes back.

Former world No. 5 and current Korean national team coach Joo Sae-hyuk (주세혁) training with Fastpong