Unimate: A programmable arm joins the line
Unimation · General Motors
Unimation's Unimate, a teach-and-replay hydraulic arm, starts unloading die castings at a GM plant in New Jersey. It made the reprogrammable arm a product, seeding an industry with 5 million factory robots in use by 2025.
A die-casting job nobody wanted
In 1961 a hydraulic arm called the Unimate began unloading hot castings from a die-casting machine at a General Motors plant in New Jersey. The Henry Ford, which holds the machine as object 76.115.1, records that it was installed at a General Motors plant in Trenton in 1961 to unload a die-casting press, and identifies it as the first Unimate used on an assembly line. The arm came from Unimation, the company formed by the inventor George Devol and the engineer Joseph Engelberger, and it carried out a fixed sequence of motions that a person had recorded beforehand. The plant made small metal hardware for car interiors, such as door and window handles, gearshift knobs and light fixtures, according to the International Federation of Robotics chronology of the industry, and the robot's task was to lift the castings out of the machine and stack them.
The narrow claim worth defending is that this was the first time a general-purpose, reprogrammable manipulator was sold to a manufacturer and kept at work on a production line. IEEE Spectrum's historical account calls it the first successful application of an industrial robot, a phrasing more careful than "first robot" and a better match for the evidence. The broader judgment, that the 1961 installation opened an industry that now runs millions of machines, rests on what came later. The machine itself was slow to sell, heavy, and unable to sense anything about the part it was holding. This report keeps those levels apart, and it flags where the record disagrees about the year, the town, and even which surviving machine is the first one.
Hard automation and the case for a general-purpose arm
By the 1950s American factories were heavily mechanized, but most of that machinery was built for one job. A transfer machine that moved a part from one station to the next did so through cams, limit switches and linkages shaped to that part and that motion, and when the product changed the machine had to be rebuilt. Devol's patent application, filed on 10 December 1954, starts from exactly this problem. The text of US patent 2,988,237, Programmed article transfer observes that cam control was common but so costly and specialized that only large-volume tasks were worth mechanizing, while manual handling was flexible but wasteful of labor. Devol proposed what the patent calls "a more or less general purpose machine", one whose task could be changed by changing its program rather than its hardware.
The patent is candid about its debts. It notes that programming techniques had long been used in weaving and knitting machines and in automatic machine tools, and places its novelty in bringing flexible programming to the routine job of carrying articles from a supply point to a delivery point. Devol's background explains his choice of tools. The National Inventors Hall of Fame says he began to see the value of factory automation while working on magnetic recording in the 1930s, and a 2011 IEEE Spectrum profile lists an automatic door sold as the Phantom Doorman, radar work at Sperry Gyroscope during the Second World War, and digital magnetic recording among his earlier projects. The robot was, in a real sense, a recording device given an arm.
What the 1954 patent actually describes
The machine in Devol's drawings is less humanlike than the Unimate that later became famous. Its worked example is a carriage running on rails beside a row of pallets, with a transfer head pushed horizontally by a telescoping hydraulic arm and raised and lowered on an elevator, plus a gripper that opens and closes. The patent says swiveling motions could be added, but the illustrated machine moves cartons from pallets to a conveyor with straight-line strokes.
The control scheme is the real content. Each motion has its own track on a magnetic drum, which stores a sequence of target positions as binary codes. A sensing head travels with the transfer head along a stack of notched strips, each strip encoding a numbered position. The machine keeps moving until the code read from the drum matches the code read from the strips, a comparison made by magnetic coincidence detectors, and an anticipating sensor slows the head as it nears its target so it can travel fast and still stop accurately. The drum then steps to the next instruction. Programming needed no language at all: the patent explains that the apparatus could be cycled under manual control, with each critical position written onto the drum for later playback, and it prefers magnetic storage because it can be erased and re-recorded.
The production Unimate kept that teach-and-repeat principle while changing the body. IEEE Spectrum reports that a team of six built the first arm with five degrees of freedom before moving to six axes that imitated a shoulder, elbow and wrist, driven by a hydraulic supply at 1,000 pounds per square inch, and that the prototype weighed about 1,360 kilograms and could lift parts of up to 45 kilograms. The IFR chronology adds that the early machines stored the angles of their joints during a teaching phase and replayed them in operation. Some later summaries, including the Hall of Fame's, credit the 1954 filing itself with a six-axis arm. The patent text is more modest, and the gap is a reminder that retrospective accounts tend to read the finished product back into the founding document.
From patent to purchase order
The inventor needed a seller. Devol met Engelberger at a cocktail party in 1956, two years after filing, and according to IEEE Spectrum the two talked about Isaac Asimov's robot stories before turning to business. Unimation, its name a blend of "universal" and "automation", was formed soon afterward. Engelberger persuaded Norman Schafler, chief executive of Condec Corporation, the parent of a small controls firm Engelberger had set up, to fund development. Condec bought Unimation in 1960 and acquired Devol's patent when it was finally issued, on 13 June 1961.
Engelberger then went looking for customers. Spectrum reports that he visited 15 automobile factories and 20 other manufacturers to learn what an arm could usefully do. General Motors bought first, paying 18,000 dollars for a machine that cost an estimated 65,000 dollars to build, a loss Engelberger accepted in exchange for being able to name GM as a client. The Hall of Fame's account of the invention says the Unimate 1900 was shown to the public at a trade show in Chicago in 1961 and that the first arm stacked hot metal at the GM plant. A trade-press retrospective in Assembly magazine gives more detail: the robot dropped hot handles and knobs into a cooling bath, took about an hour to program, and fed a line where people did the trimming and buffing by hand.
The business case was not obvious. Spectrum notes that the list price had reached 35,000 dollars by the early 1970s and that Unimation did not turn a profit until 1975, 19 years after its founding. Engelberger worked the publicity side hard. In 1966 the arm appeared on The Tonight Show with Johnny Carson, where it opened a can of beer, conducted the band and knocked a golf ball into a cup, according to the Hall of Fame and Spectrum. Spectrum reads the appearance as a deliberate effort to make a machine that workers might fear look friendly.
Which year, which town, which machine
Readers will meet several versions of this event, and they do not agree. IEEE Spectrum's robots guide says Devol and Engelberger built a prototype in 1958. The IFR places the development of the first robot in 1959, and an engineering.com obituary of Engelberger says the prototype, Unimate number 001, was completed and installed at a GM die-casting plant that same year. The museum record, Spectrum, the Hall of Fame, the Computer History Museum and the IFR's own entry for the GM installation all give 1961 for the robot entering production work, the same year the patent issued. Chronicle dates the event to 1961 because that year has the strongest documentary support for a machine at work on a production line, and because a prototype or trial is a different claim from a sale. The 1959 account cannot be ruled out, but it rests on later recollection rather than a record of its own.
The place is a smaller puzzle. The Henry Ford, Spectrum and the IFR say Trenton, while the Hall of Fame blog and the Robots guide say Ewing Township. Assembly magazine puts GM's Ternstedt division plant in Ewing, a township on the edge of Trenton, which suggests one building described by its postal city in some sources and its municipality in others.
Even the object is contested. Spectrum and The Henry Ford place the first installed Unimate in Dearborn, Michigan, while the Hall of Fame says the original resides in the Smithsonian, and Assembly says the Smithsonian's machine ran nearly around the clock for more than a decade. Both institutions may hold early units, but the word "first" cannot belong to both. The Unimate also had a near-contemporary rival. The IFR records that American Machine and Foundry installed six Versatran robots, a cylindrical design whose name came from "versatile transfer", at a Ford factory in Canton in 1962, and Assembly says AMF developed it a year before the Unimate debut. The Unimate's priority is real but narrow.
A robot with no sense of the world
For a history of artificial intelligence, the most important fact about the Unimate is what it lacked. It had encoders on its own joints but nothing that reported on the world: no vision, no touch, no way to notice that a casting was missing or out of place. It repeated a recorded path exactly, which suited a die-casting cell where the part always arrived in the same spot and failed anywhere that varied. Its program was a list of positions, not a model of a task.
The contrast with AI laboratories in the same years is sharp. At MIT, Heinrich Ernst built MH-1, a mechanical hand driven by the TX-0 computer, as the subject of a doctoral thesis supervised by Claude Shannon and completed in 1962. The Computer History Museum's exhibit on putting robots to work shows that hand next to the Unimate, along with a hydraulic arm built at Stanford's Artificial Intelligence Project between 1966 and 1968 by Victor Scheinman and others to test whether a manipulator could keep up with the computer controlling it. At SRI, Shakey, the subject of research from 1966 to 1972, became the first mobile robot able to perceive and reason about its surroundings (see Shakey enters the world). The laboratory machines asked how a robot could decide what to do. The Unimate assumed someone had already decided.
The two lines met in the 1970s. The IFR records that Scheinman formed Vicarm in 1973 to sell an industrial version of the Stanford arm, which had used touch and pressure sensors for small-parts assembly; the new arm was controlled by a minicomputer. Unimation bought Vicarm and with it the VAL programming language, according to Spectrum, and the Robots guide dates the resulting PUMA arm to 1978. The factory robot thus gained a computer and a language from the AI laboratory, while the laboratory gained a commercial market that justified building arms at all. Teaching by guiding the arm through its motions never went away: Kawasaki's history of its Unimate line still describes industrial robots as machines whose movements are configured through teaching and programming and then repeated for hours.
Welding lines, Japan and the labor question
Adoption was slow at home and faster abroad. Spectrum reports that GM installed its first two spot-welding robots at Norwood, Ohio, in 1967 and a line of 28 robots at Lordstown, Ohio, in 1970. The IFR dates the first Lordstown spot-welding robots to 1969 and says they let more than 90 percent of body welding be automated, against 20 to 40 percent at conventional plants. Lordstown then became a byword for conflict over automation. Spectrum reports that a 1971 retooling led to layoffs of at least 700 workers and raised line speed from 60 to 101 cars per hour, that workers filed more than 16,000 grievances, and that a strike shut the plant for 22 days in March 1972, although the robots were only one part of a broader push for output. In Europe, the IFR records the first industrial robot sale in 1967, a Unimate installed at Metallverken in Sweden.
Japan proved the larger market. After Engelberger held a long question-and-answer session with Japanese industrialists in 1966, Kawasaki won the right to build the machine. Kawasaki Heavy Industries' corporate history records a technical agreement with Unimation in 1968 and commercial production in 1969 of the Kawasaki-Unimate 2000, the first industrial robot made in Japan. The engineering.com obituary dates the license to 1966, which may conflate the talks with the agreement. Kawasaki's own account describes the first model as a 1.6-ton machine with a 12-kilogram payload, sold into a period of labor shortage, and says its redesigned 2630 type added a sixth degree of freedom and became a best seller in car factories for years. By the mid-1980s, Spectrum notes, Japan had nearly 70 percent of the world's robots. Westinghouse bought Unimation in 1983 for 107 million dollars, according to the Robots guide, and later sold it to Staubli.
The scale of what followed is documented by the industry's own statistics. The IFR's World Robotics 2026 release puts the operational stock of industrial robots at a record 5 million units in 2025, after more than 600,000 installations that year. The consequences for workers remain debated. Acemoglu and Restrepo's study of US labor markets estimates that, for the rise in robot use between 1990 and 2007, one more robot per thousand workers reduced the employment-to-population ratio by about 0.18 to 0.34 percentage points and wages by 0.25 to 0.5 percent, an effect the authors separate from trade and offshoring. Those numbers describe the Unimate's descendants, not the machine in New Jersey, but they are the downstream question it raised.
The honest limits
Several things the Unimate did not demonstrate deserve plain statement. It did not show that machines could adapt, perceive or learn; it showed that a recorded sequence of motions could be replayed reliably in a hot, dirty job. It was not a commercial success in its first years: the first sale was made at a large loss and the company needed 19 years to reach profit. Early use was concentrated in a few tasks in the car industry, chiefly die casting and spot welding, and even in Japan, Spectrum reports, the robot took a decade to be widely adopted.
The evidence base has its own weaknesses. Almost every detailed account of 1958 to 1961 was written decades later, often by companies, trade bodies and halls of fame with an interest in a clean origin story, and they disagree about dates, places, prices and weights. Spectrum gives the prototype at about 1,360 kilograms, Assembly describes a 4,000-pound arm, and the IFR says two tons. The claim to be first depends on definitions, since AMF's Versatran was developed in parallel and reached a Ford plant within a year. What survives scrutiny is a patent, a sale, a surviving machine and a growing industry, not a single dramatic unveiling.
How to place Unimate in AI history
The Unimate belongs in a history of AI as the body that arrived before the mind. In 1956, the year the Dartmouth proposal gave artificial intelligence its name, Devol and Engelberger formed the company that would put a programmable arm into a factory. The two projects shared a year and almost nothing else. AI researchers spent the following decades trying to give machines perception and planning, while industrial robotics grew into a large business on machines that had neither, by reshaping the world to fit the robot: fixtures, fences and parts presented in the same place every time.
That division is the useful lesson. The Unimate showed that a general-purpose machine could pay its way long before it was intelligent, provided its surroundings were engineered to be predictable. Much of later robotics, from the Stanford arm to machine vision on the factory floor to vehicles that drive in open terrain, can be read as an effort to relax that constraint. The patent's lasting idea was the program as the thing that changes between jobs, which looks obvious now and in 1954 took an inventor who had spent years recording signals onto magnetic media to see. Its limit, a robot that could only repeat what it was shown, marks the gap that robot perception and learning have been working to close ever since.
References
- Devol, George C. (1961). Programmed article transfer, US patent 2,988,237.
- The Henry Ford (2026). Robot, First Unimate Robot Ever Installed on an Assembly Line, 1961.
- Marsh, IEEE Spectrum (2022). In 1961, the First Robot Arm Punched In.
- Malone, IEEE Spectrum (2011). George Devol: A Life Devoted to Invention, and Robots.
- IEEE Spectrum (2026). Unimate, ROBOTS: Your Guide to the World of Robotics.
- National Inventors Hall of Fame (2026). George Devol: Industrial Robot, U.S. Patent No. 2,988,237.
- National Inventors Hall of Fame (2019). The Invention of the Industrial Robot.
- International Federation of Robotics (2026). Robot history.
- International Federation of Robotics (2026). Five Million Robots now Operate in Factories Globally.
- Wright, engineering.com (2015). Robot pioneer Joseph F. Engelberger Dies at Age 90.
- Weber, Assembly (2011). Happy Big 5-0 Robot.
- Computer History Museum (2026). Putting Robots to Work.
- Ernst, MIT (1962). MH-1, a computer-operated mechanical hand.
- SRI International (2026). Shakey the Robot.
- Kawasaki Robotics (2022). The Story of the Kawasaki-Unimate: Japan's First Domestically Manufactured Industrial Robot.
- Kawasaki Heavy Industries (2026). History of the robot business.
- Acemoglu and Restrepo, NBER (2017). Robots and Jobs: Evidence from US Labor Markets.
Cite this report
Chronicle. (2026, October 18). Unimate (1961): A programmable arm joins the line. https://epoch.founden.co/events/unimate
Connected developments
Follow the ideas behind this event and the developments that came after it. Connections indicate conceptual or historical relationships, not proof of sole causation.
AI named the year Unimation formed · 1956Dartmouth workshop: Artificial intelligence gets a nameAI lab robot that could perceive · 1966Shakey the robot: Shakey enters the worldRobots leave the engineered cell · 2005DARPA Grand Challenge: Driverless across the desertFrom factory floor to public road · 2020Waymo fully driverless public service: The driver's seat goes empty