# History of robot arms

October 23, 2026 · https://tatbot.ai/updates/2026-10-23/

In 1961, a Unimate unloaded a die-casting press at a General Motors plant in Trenton, New Jersey. The surviving machine is now in The Henry Ford's collection. Its job gives this history a useful starting point: take a part from a particular machine and move it somewhere else, repeatedly. The factory application came before today's familiar images of robots recognizing objects or responding to spoken instructions. [The installed Unimate](https://www.thehenryford.org/collections/explore/artifact/183434)

Making an arm useful involves several problems at once. It needs enough force, suitable reach and a way to direct movement. Sensing can reveal what actually moved. Someone must also arrange the parts, attach a tool and decide what counts as a successful movement. Changes in each of those jobs helped bring arms into factories, laboratories and home workshops. Their histories do not form a single climb toward the cheapest machine or the smallest advertised error.

## Teaching a machine where to go

Remote manipulation already had a different purpose before Unimate. In a 1949 Argonne report, Raymond Goertz described a mechanical manipulator for work across a protective barrier. Gears, drums and cables carried movements from the operator's handle to a remote tong. Six motions positioned and oriented the tong; another closed it. The human supplied the continuing decisions and movement. This was manipulation at a distance without an autonomous task program. [Goertz's report, pages 3–6](https://www.osti.gov/servlets/purl/1054625)

George Devol's programmed article-transfer patent addressed a different arrangement. Filed in December 1954 and granted in June 1961, it described storing position information and using it to control a sequence of transfers. Its motivation included changing a machine's work without rebuilding the specialized mechanisms that imposed a fixed sequence. The patent date records a disclosure; the museum's 1961 installation record establishes a particular factory application. [Devol's patent](https://patents.google.com/patent/US2988237A/en)

A later Unimation brochure, dated 1962–1968 by the museum, makes the programming tangible. A setup worker moved the arm with teaching controls, recorded positions and recorded gripper actions. The machine then repeated the sequence. The brochure described hydraulic power, magnetic memory and capacity for up to 200 sequential points. These are specifications for the brochure's machine, rather than proof of every detail inside the earlier installed unit. [Contemporary brochure](https://www.thehenryford.org/collections/explore/artifact/373631)

The same brochure promised quick training and savings through reuse on different jobs. Those were the maker's sales claims. In 1981, Ayres and Miller reported that interviewed industrial users most often kept a robot specialized to one application, partly because of constraints on moving it. Changing a program and moving an arm into another job were different forms of flexibility. [Industrial-use discussion, pages 15–16](https://publications.ri.cmu.edu/storage/publications/pub_files/pub3/ayres_robert_1981_1/ayres_robert_1981_1.pdf)

Reaching the destination and controlling the route were distinct problems. A point-to-point sequence specifies destinations; continuous-path control also governs movement between them. In another patent, filed in May 1964 and granted in February 1967, Devol described combining these modes. His example moved a part through a narrow opening into a spray enclosure, then switched control modes to position it at its destination. The document discloses a control arrangement, rather than establishing when a factory bought or operated it. It also shows why controlling a path did not simply make endpoint positioning obsolete. [Combined motion controls, columns 1–2, 5–6 and Figure 1](https://patents.google.com/patent/US3306471A/en)

## Designing an arm for the computer

Hydraulics use pressurized liquid to exert force. Pneumatic mechanisms use compressed gas. Electric motors turn electrical energy into mechanical motion, often through a transmission. Each arrangement still needs a way to direct its movement. The early electric research arms make that distinction particularly clear.

Laboratory notes describe Victor Scheinman's electric Stanford arm as designed in 1969 and in use since 1970. They distinguish later versions with changed reach, capacity and stiffness. A 1973 paper describes a Stanford-derived implementation at the Jet Propulsion Laboratory (JPL). Three rotary axes intersected at one point, simplifying the calculation of joint positions for a desired hand position. The designers chose this geometry to reduce computation and develop smooth joint movements. [Stanford laboratory notes](https://www.saildart.org/NOTES%5B1,VDS%5D); [JPL design paper](https://www.ijcai.org/Proceedings/73/Papers/032.pdf)

The JPL machine used six DC torque motors. Its transmission varied with the joint: strain-wave reducers on inner rotary axes, a rack and pinion for the extending axis and spur gearing at the sixth joint. Position and speed feedback let the controller respond to actual movement. [Motor, gearing and feedback description](https://www.ijcai.org/Proceedings/73/Papers/032.pdf)

That is the useful meaning of a **servo system**: feedback participates in controlling the motion. The term does not identify one species of motor. A stepper motor can also operate with feedback. Oriental Motor's engineering explanation, for example, describes a stepper system with a resolver that monitors the rotor. Motor construction and control arrangement are separate questions. [Stepper feedback example](https://www.orientalmotor.com/stepper-motors/technology/stepper-motor-basics.html)

The same Stanford notes describe a separate hydraulic arm with servo valves and potentiometer feedback. Feedback control did not belong exclusively to electric motors. Within electric drives, brushed and brushless describe how current is switched as the rotor turns. Brushes make sliding electrical contact with a commutator; a brushless drive switches the windings electronically using rotor-position information. Removing brushes removes that wear mechanism, while requiring the matching electronics. It does not specify the arm's gearing or task accuracy. [Hydraulic servo description](https://www.saildart.org/NOTES%5B1,VDS%5D); [Brushed commutation](https://support.maxongroup.com/hc/de/article_attachments/4415193502226); [Brushless drive](https://support.maxongroup.com/hc/en-us/article_attachments/4415181729042)

A 1974 JPL report by Richard Lewis separated planning from real-time execution. Its example traced a task through joint trajectories, nominal torques, feedback corrections and motor currents. Yet its conclusion said real-time software and integration were still under development, with sensor-responsive planning to follow. The separation explains what computer control involved, while the unfinished work prevents treating the architecture diagram as proof that the complete autonomous system already worked. [Software architecture and status, pages 39–42](https://ntrs.nasa.gov/api/citations/19740011910/downloads/19740011910.pdf)

Research and commercial machines also overlapped. A 1981 CMU report's chronology records General Motors' PUMA specification in 1977, Unimation's acquisition of Scheinman's Vicarm and a prototype shipment in 1978. It is a near-contemporary account of that transition, rather than a specification for every later PUMA model. Programmable electric manipulation was becoming a product as well as a laboratory instrument. [Ayres and Miller, chronology page 48](https://publications.ri.cmu.edu/storage/publications/pub_files/pub3/ayres_robert_1981_1/ayres_robert_1981_1.pdf)

## An assembly arm could benefit from giving way

Assembly exposed a weakness in the idea that an arm should resist movement equally in every direction. A peg approaching a chamfered hole may need a little sideways movement to align itself while resisting movement along the insertion direction.

Hiroshi Makino's team at Japan's University of Yamanashi developed SCARA with industrial partners. The name meant Selective Compliance Assembly Robot Arm. His history distinguishes the October 1978 prototype, a second in May 1980 and commercial variants in 1981. [Makino's history](https://www.jstage.jst.go.jp/article/jrobomech/26/1/26_5/_pdf/-char/en)

The mechanisms were mixed. The first two axes used DC motors, optical encoders and strain-wave reduction. Tool rotation used a stepper with belts. The 1982 report describes pneumatic vertical motion for Prototype II. Vibration in the first prototype led to changes including velocity feedback in the second. There was no single motor upgrade that accounted for the whole machine. [Prototype II, 1982, page 379](https://www.jstage.jst.go.jp/article/jjspe1933/48/3/48_3_378/_pdf#page=2)

Early control used an NEC TK-80 microcomputer to shape acceleration and deceleration in software, adapting motion ideas from fixed cams to changeable movements. Makino's insertion illustration also shows why lateral yielding can help while tilting causes trouble. Useful compliance depends on its direction. [Construction and motion control](https://www.jstage.jst.go.jp/article/jrobomech/26/1/26_5/_pdf/-char/en)

## What a gearbox gives and takes

A motor that turns quickly with modest torque can drive a slower, stronger joint through reduction. That helps package an arm, but the transmission contributes its own behavior. The JPL designers discussed lost motion, elastic windup and the way reduction magnifies the motor's apparent inertia at the output. A controller's view of the motor is therefore not automatically a complete measurement of the tool. [JPL transmission discussion](https://www.ijcai.org/Proceedings/73/Papers/032.pdf)

Different reducers achieve that trade in different ways. A strain-wave drive uses an elliptical wave generator to deform a flexible toothed cup inside a circular toothed ring. A small difference in tooth counts produces a large reduction. Nabtesco's RV mechanism combines an initial gear stage with eccentric motion of gears against surrounding pins. These are particular mechanisms, rather than interchangeable names for all robot gearboxes. [Strain-wave operation](https://legacy.harmonicdrive.net/reference/operatingprinciples/); [RV operation](https://www.nabtescoprecision.com/engineering/)

Backlash is lost movement when clearances must be taken up after reversal. Elastic deformation is different: parts can twist or bend while transmitting force. A reducer promoted as having zero backlash does not thereby make the complete arm infinitely stiff or its tool perfectly located. The useful question is what the whole mechanism does under the load and movement the task requires.

Sensor placement changes what the controller can observe. A motor-side encoder measures before the gearbox; an output-side encoder measures the joint after it. HEIDENHAIN supplies both arrangements, including combined devices. The secondary-encoder study modeled that flexible transmission when designing its controller. Measuring the joint more directly can expose error hidden by a motor reading, but it does not measure every bend in the structure or locate the tool in an external coordinate system. [Encoder arrangements](https://www.heidenhain.com/industries/robotics); [Secondary-encoder control study](https://arxiv.org/pdf/2101.04992)

A contemporary small arm shows why broad labels can mislead. UFACTORY describes the Lite 6 using direct-drive and harmonic-drive language. Its support table lists reduction ratios of 100 for the first three joints, then 1, 2.4 and 1 for the wrist joints. Calling the entire arm direct drive would conceal that mixture. [Lite 6 description](https://www.ufactory.cc/product-page/ufactory-lite-6/); [Joint ratios](https://docs.supportarticle.ufactory.cc/support_articles/hardware/ufactory-joint-parameters.html)

## Removing gears did not remove the engineering

Carnegie Mellon's direct-drive research tested what happened when motors drove joints without reduction gear trains. Eliminating those trains removed important sources of friction and lost motion, but the motors had to provide the joint torque themselves.

The 1981 design paper described a chain of loads: a wrist motor's mass loaded the elbow, and both loaded the shoulder. Bigger motors near the tip could therefore demand bigger motors farther back. Their layout made one motor serve as a counterweight, while heat dissipation remained further work. Removing gears made motor choice, placement and cooling more consequential. [Direct-drive design, pages 2 and 7–8](https://publications.ri.cmu.edu/storage/publications/pub_files/pub4/asada_haruhiko_1981_1/asada_haruhiko_1981_1.pdf)

The second CMU arm, described in a 1985 report by Takeo Kanade and Donald Schmitz, used samarium-cobalt brushless DC motors, resolvers and current feedback. Lessons from the first arm included inadequate motor torque, heavy construction and limited control computation. The new design combined mechanical changes with a multiprocessor control system. Direct drive was a system project. [CMU Direct-Drive Arm II](https://www.ri.cmu.edu/pub_files/pub3/kanade_takeo_1985_2/kanade_takeo_1985_2.pdf)

That report contains an inviting specification table, including a 900 mm workspace radius and an accuracy figure of ±0.1 mm. Its conclusion explicitly calls the table expected performance and describes experiments still to come. A design target is not a measured result. Reading the surrounding text changes what the numbers can support.

Actual experiments on the earlier arm offer a more bounded result. A 1982 paper tested repeated movements between the same starting and destination positions, then examined settled joint angles. Model-based compensation changed both average error and scatter. The experiment measured particular joint behavior; it did not establish externally measured tool accuracy everywhere the arm could reach. [Direct-drive control experiments](https://www.ri.cmu.edu/pub_files/pub3/asada_haruhiko_1982_1/asada_haruhiko_1982_1.pdf)

JPL's 1973 analysis explained why computing motor torque did not remove feedback: models were imperfect and parameters changed. It compared computed torque with a conventional position servo without declaring a universal winner. Computation and further analysis mattered. [Control comparison, pages 1 and 37–40](https://ntrs.nasa.gov/api/citations/19730012471/downloads/19730012471.pdf)

## A small number needs a measurement question

An arm can return closely to the same place while consistently missing the coordinate a program intended. **Repeatability** describes the grouping of repeated results. **Absolute positioning accuracy** asks how closely the result matches an independently specified target. **Path accuracy** concerns the route taken between points. Orientation error matters too: a tool can have its tip near the right location while pointing the wrong way.

ABB's specification for the IRB 1100-4/0.58 makes the difference concrete. This is a six-axis industrial arm with 580 mm reach and a 4 kg rated payload. Revision V gives 0.01 mm pose repeatability and 1.18 mm path accuracy. They are different tests of the same model, rather than competing descriptions of one universal precision. The path conditions include maximum rated load and offset, 1.6 m/s motion and an inclined test plane with all six axes moving. [ABB specification, section 2.7](https://library.e.abb.com/public/59c58801f17f4f509a97c53f36b5cc20/3HAC064993%20PS%20IRB%201100-en.pdf)

The table's pose-accuracy footnote is equally consequential. It compares a manually taught position with the mean of executed positions. That does not promise that an arbitrary coordinate imported from a drawing will be reached within 0.01 mm. Teaching a point on the installed machine and calculating one from an external model create different demands.

ISO 9283's available standard preview explicitly requires matching conditions for comparisons between robots, including test volume, loads, velocities, paths, cycles and environment. Preparation and measurement provisions address warm-up, thermal stability and calibrated measuring equipment. A seller's isolated repeatability number leaves much of that story untold. [ISO 9283 preview, scope and test conditions](https://preview.sist.si/sist-preview/22244/d058a9d8bacf41cd85def67ab70afc87/ISO-9283-1998.pdf)

Resolution asks another question: how small a change can the command or sensor distinguish? Feetech's 2020 STS3215 datasheet separately lists 0.088° electronic resolution and gearbox backlash of up to 0.5°. Those are component specifications, not a measurement of an assembled arm's tool. A small sensor increment cannot by itself establish how closely a gripper will place an object. [STS3215 datasheet, page 4](https://www.feetechrc.com/Data/feetechrc/upload/file/20200611/6372749961523760249976542.pdf)

Modern programming interfaces preserve the choice. In Universal Robots' SW5_25 manual, MoveJ coordinates joints to arrive together; the tool can follow a curved route. MoveL specifies a straight tool-center-point path. MoveP adds constant-speed movement with blends for processes such as gluing or dispensing. Teaching the same endpoints can therefore leave important parts of the task unspecified. The motion type and speed affect what happens between them. [Motion commands](https://www.universal-robots.com/manuals/EN/HTML/SW5_25/Content/prod-usr-man/complianceUR5e/SW_sections/first_program/move.htm)

## Calibration connects the arm to the work

Joint sensors describe the mechanism through its own coordinates. A task also needs relationships among the robot base, the workpiece and the tool. The tool-center point identifies the working point relative to its mount; changing the tool changes that relationship. Locating a part with a camera adds another coordinate relationship. Errors in these relationships can remain even when repeated joint readings look excellent. [Tool-center-point definition](https://www.universal-robots.com/manuals/EN/HTML/SW5_24/Content/prod-usr-man/software/PolyScope/content/installation_g5/installation_TCP_configuration_en.htm); [NIST on calibration and registration](https://www.nist.gov/el/intelligent-systems-division-73500/calibration-and-registration-tools)

This problem was already visible in 1986. Henry Stone, Arthur Sanderson and Charles Neuman described identifying the individual geometry of seven PUMA 560 arms with external ultrasonic measurements. They evaluated two in detail on a three-dimensional grid using a separate measuring jig and an optical-encoder-equipped milling machine. The work connected a manufactured arm's geometry to its controller's model and checked the result with another instrument. [Arm Signature Identification](https://publications.ri.cmu.edu/storage/publications/pub_files/pub4/stone_h_w_1986_1/stone_h_w_1986_1.pdf)

In a 2019 study, Karl Van Wyk, Joe Falco and Geraldine Cheok tested calibration using a seven-axis KUKA arm and independent measurement equipment. They adjusted joint zero offsets and checked performance at held-out positions. Their method improved positioning, but residual errors remained. The paper separates measured validation from estimates of improvement across the full workspace. These are results for that setup. [Calibration study](https://arxiv.org/pdf/1908.07273)

Calibration can correct a repeatable model error. It cannot turn every changing load, temperature or contact condition into a fixed offset.

Contact work adds another control problem. Matthew Mason's 1981 paper distinguished passive compliance built into a mechanism from active compliance implemented in the servo loop. A controller can regulate force or the relationship between force and motion. Impedance control specifies a chosen mechanical response, such as yielding like a spring. [Mason's paper abstract](https://publications.ri.cmu.edu/compliance-and-force-control-for-computer-controlled-manipulators); [MIT manipulation control notes](https://manipulation.mit.edu/force.html)

A concrete mechanical example appears in Paul Watson's remote-center compliance patent, filed in 1976 and granted in 1978. Flexible links let a held part make small sideways adjustments and rotate around a point near its tip under insertion forces. The disclosed arrangement illustrates why useful yielding depends on its location as well as its direction. [Remote-center mechanism, Figures 1–5](https://patents.google.com/patent/US4098001A/en)

Force feedback also brings limits. In 1987, Steven Eppinger and Warren Seering analyzed how flexibility between an actuator and a force sensor could make a controller unstable as its reaction to force error increased. Their study used simplified models. Adding a sensor did not remove the need to understand the structure between sensing and actuation. [Force-control analysis](https://www.cs.cmu.edu/~cga/force-control/eppinger2.pdf)

## Access changed through components and integration

A December 1985 Dataquest report on U.S. semiconductor robotics listed approximate dollar prices of $48,000 for a clean-room PUMA 560 with VAL-Plus control and $53,000 with VAL-II. The table does not identify new-versus-used status or itemize taxes, shipping and installation. Its discussion separates robots from process equipment, controls, software and integration. These are period analyst estimates, not retrieved invoices. Comparing them with today's component subtotal would mix dates, configurations and tasks. [Dataquest, pages 5 and 7](https://archive.computerhistory.org/resources/access/text/2013/04/102723443-05-01-acc.pdf#page=266)

Universal Robots' 2018 e-Series launch illustrates access through a packaged platform. The company combined built-in force/torque sensing with its programming interface and developer tools. The SW5_25 UR5e manual specifies 850 mm reach, a 5 kg payload and pose repeatability of ±0.03 mm under its ISO 9283 reference. That number remains a manufacturer specification, rather than a comparison with the ABB path test. [e-Series launch](https://www.universal-robots.com/news-and-media/news-center/universal-robots-launches-e-series-setting-a-new-standard-for-collaborative-automation-platforms/); [UR5e manual](https://www.universal-robots.com/manuals/EN/HTML/SW5_25/Content/prod-usr-man/complianceUR5e/H_g5_sections/appendix_g5/tech_spec_sheet.htm)

Model names need dates too. Universal Robots now describes the former UR5e as UR7e-920. Its FAQ says the arm is mechanically unchanged and the new reach figure uses a revised measurement method. Here, a larger reported reach accompanies a new way of measuring the same mechanism. [Current naming explanation](https://www.universal-robots.com/insights/faq/)

At a smaller scale, UFACTORY's Lite 6 page advertised a starting price of **US$2,999**, checked October 22, 2026. It lists 440 mm reach and 600 g payload. Its advertised repeatability is 0.5 mm, while its V2.3.0 manual lists ±1 mm; matching test conditions are missing. Its store lists an arm with integrated controller and Ethernet cable, shipping EXW from China with import tariffs assigned to the buyer. The starting figure is an advertised new-product price, not a verified tax-and-shipping-inclusive checkout total. [Lite 6 specifications and price](https://www.ufactory.cc/lite-6-collaborative-robot/); [Manual, page 146](https://www.ufactory.cc/wp-content/uploads/2024/01/Lite6-User-Manual-V2.3.0.pdf#page=146); [Store bundle and shipping terms](https://www.ufactory.cc/product-page/ufactory-lite-6/)

The SO-101 project takes another route. Its follower has five arm joints and a gripper, using six geared STS3215 servo units. Printable structure, build directions and open software let a builder assemble a teaching platform from reusable components. The servos combine gearing, angle sensing and local control; the builder does not have to design each of those functions from scratch. [LeRobot assembly documentation](https://huggingface.co/docs/lerobot/so101); [Servo specification](https://www.feetechrc.com/Data/feetechrc/upload/file/20200611/6372749961523760249976542.pdf)

Its hand-moved leader deliberately uses different gearing: 345:1 at the shoulder lift, 191:1 at the base and elbow, and 147:1 at the wrists and gripper. LeRobot explains the mix as a compromise between supporting the arm's weight and moving it easily. Backdriving means moving a joint from its output so that the transmission turns the motor. A teaching handle and a powered follower place different demands on the same kind of component. [Leader gearing](https://huggingface.co/docs/lerobot/so101?assembly=Leader)

The project's pinned parts list gives a US-column subtotal of **$121.94 for one follower**, or **$229.88 for a leader-and-follower pair**, read October 22, 2026. These are project estimates for listed components, not independently verified seller quotations. Printed parts, cameras, a computer, assembly labor, taxes and shipping are outside those subtotals. Comparing $121.94 directly with a complete integrated arm would hide much of what each figure buys. [Pinned project bill of materials](https://github.com/TheRobotStudio/SO-ARM100/blob/a758567c3978dfeefe282ede0500085a48fe8f78/README.md)

Keeping a working installation useful adds another kind of cost. Universal Robots' service handbook specifies inspections and says that replacing a joint invalidates the existing calibration. Its recommended recalibration process uses another robot and dedicated tooling. A replaceable module can simplify a repair, while recovering the relationship between the repaired mechanism and an existing program still takes equipment and work. The handbook documents that requirement; it does not supply a universal maintenance bill. [e-Series service handbook, sections 4 and 5.2.12](https://www.universal-robots.com/manuals/EN/PDF/SW5_24/service-manual_e-series/e-Series_Service_Manual_en.pdf)

## Learning changes how a task is specified

The 2023 ALOHA project used low-cost arms to learn manipulation from human demonstrations: 50 per real-world task, except 100 for threading Velcro. The authors report about 30–60 minutes of collection per task, including resets and operator mistakes. A policy predicted joint-position targets from observations. Local feedback controllers inside the Dynamixel motors tracked those targets. Learning changed how movements were chosen; motor controllers still used feedback to execute them. [ALOHA data collection and control](https://tonyzhaozh.github.io/aloha/aloha.pdf)

ALOHA's main real-world tests varied object positions along a 15 cm reference line and evaluated policies in 25 trials each. Results differed by task. The authors also described limits from motor torque and perception. These task outcomes do not establish absolute tool accuracy, service life or a complete installation cost. [Evaluation and limitations, section V and Appendix F](https://tonyzhaozh.github.io/aloha/aloha.pdf)

The costs around the arm were visible long before learning systems. Writing in 1981, Ayres and Miller argued that setup and programming could dominate one-off applications, while dedicated machinery could outperform then-current robots in long, fixed production runs. Their forecasts belong to their period, but their comparison identifies a durable variable: the amount of change the application needs. [Manufacturing analysis, pages 10–14](https://publications.ri.cmu.edu/storage/publications/pub_files/pub3/ayres_robert_1981_1/ayres_robert_1981_1.pdf)

A die-casting transfer, a compliant assembly move and a desktop demonstration each organize work differently. The arm's value comes from fitting its mechanics, measurements and programming to that work. Across this history, people kept deciding what to teach, what to sense, what to calibrate and what the surrounding equipment had to provide. Those decisions help explain both the reach of inexpensive arms and the demands that an arm's headline price or repeatability cannot settle.