Industries / Industrial equipment and services / Industrial services / Welding and fabrication
Robot welding
Robot welding is welding done by machines that move the electrodes or the torch themselves, from fixed welding machines to robot arms that find the joint with a camera.
6
Milestones achieved, from 1886 to 2025.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1886
Current makes the weld
Elihu Thomson of Lynn, Massachusetts, was granted a patent for an "Apparatus for Electric Welding" on 10 August 1886. The machine clamps two pieces of metal end to end under light pressure and sends a heavy current through them. The resistance of the joint heats the ends until they fuse across their whole section. This is resistance welding, the family that also includes spot welding, which joins sheets at small points and was the job of the first welding robots.
Credit
Elihu Thomson
1936
Welding under a blanket
Lloyd Jones, Harry Kennedy and Maynard Rotermund of Union Carbide and Carbon patented a way to weld with the end of the wire buried in a layer of mineral welding composition laid over the joint, so the weld forms under a protective covering. The patent, filed in October 1935 and granted on 9 June 1936, says the wire is fed and moved along the seam "preferably by a mechanical device", and that the method avoids the trouble automatic machines had in restarting an open arc. The process is now called submerged arc welding, and its patent was written with machine welding in mind.
Credit
Lloyd T. Jones, Harry E. Kennedy and Maynard A. Rotermund, Union Carbide and Carbon
1970
A line of welding robots
General Motors installed its first two spot-welding robots, Unimates made by Unimation, at its plant in Norwood, Ohio, in 1967. In 1970 it started a line of 28 spot-welding robots at Lordstown, Ohio, where it built the Chevrolet Vega. A 1971 retooling brought layoffs and raised the line speed from 60 to 101 cars an hour. Workers filed more than 16,000 grievances, and in March 1972 a strike shut the plant for 22 days.
LessonRobots that arrive with faster lines and job cuts can provoke a revolt, and IEEE Spectrum says Lordstown hurt Unimation's reputation in its home country.
Credit
General Motors and Unimation, as told by Allison Marsh in IEEE Spectrum (2022)
1977
Electric robots take up arcs
Arc welding melts metal with an electric arc while the torch moves along a seam. Yaskawa Electric showed its first MOTOMAN robot in 1974 and calls it the first electrically driven industrial robot in Japan, but it did not sell. Yaskawa then built the MOTOMAN-L10, a jointed arm modeled on an electric robot announced by ASEA (now ABB) of Sweden and aimed at arc welding, and in 1977 sold the first one to the Nakatsu plant of the car-parts maker Yorozu. Yaskawa says it became the world's number one robot maker by sales volume in fiscal 2002.
LessonThe first robot found no buyers; the one that sold copied the layout of a rival's electric arm and aimed at one job.
Credit
Yaskawa Electric, after ASEA's design; Yorozu as the first buyer
1985
Robots see the seam
A robot that replays a taught path misses the joint when parts vary. On pressed steel sheet 1 to 2 mm thick, worn fixtures, springback and heat distortion moved the seam by as much as ±3.0 mm from where the torch expected it. Clocksin of the University of Cambridge, Bromley, Davey and Vidler of Meta Machines Ltd in Abingdon, and Morgan of the University of Oxford described a MIG welding robot (metal inert gas, a process that feeds wire through the torch) that used a camera and models of lap, T and butt joints to find the seam. It corrected the torch to within ±0.5 mm and changed its welding settings as the gap width changed.
Credit
Clocksin, Bromley, Davey, Vidler and Morgan; Meta Machines
2025
Cobots weld in small shops
A collaborative robot, or cobot, is a light robot arm built to work beside people without a safety cage. ASSEMBLY magazine reported in July 2025 that welding is one of the fastest-growing uses of cobots, and that small companies and contract manufacturers buy them for high-mix, low-volume work because they cost less and are easier to program than caged welding cells. Integrators such as Vectis Automation pair Universal Robots arms with welders from Fronius, Hypertherm, Lincoln Electric and Miller, and Fanuc said MIG welding is "by far" the most common job for its CRX cobots. Not every robot maker agrees: Kuka told the magazine it does not recommend cobots for welding, because traditional robots are faster, more repeatable and more robust.
Credit
Universal Robots, Fanuc, ABB and the integrators that package their arms for welding, as reported by Austin Weber in ASSEMBLY
Watch for
Cells that plan their welds
Path Robotics says its welding model, Obsidian, trained on what it calls millions of welds, finds weld joints, plans and makes the welds, and adapts to parts that do not fit perfectly, with no robot programming. Its website claims a first-pass yield, the share of welds accepted without rework, of "97%+". A May 2025 Machine Design article gave a figure of near 99%, against about 60 to 70% for a human welder, with no method. In 2026 Path also put the model on a four-legged robot it calls Rove, for parts too big to bring to a cell. In October 2026 it raised $300 million, said its annual recurring revenue had grown 17-fold in 18 months, and named mid-sized customers including Millerbernd Manufacturing and TYCROP.
Achieved whena customer or another outside party publishes first-pass yield, the inspection standard used, and how many of the parts were new to the system.
path-robotics.commachinedesign.complantautomation-technology.com
Claimed by
Path Robotics, May 2025 to October 2026
A robot checks its welds
Yaskawa Electric and GAZOO Racing, Toyota's motorsports division, say a MOTOMAN NEXT robot, which carries its own graphics processor for AI, has welded go-kart frames at GR's Gamagori workshop since September 2026. The robot reads each pipe frame's fit-up and joint gaps, sets the welding conditions, then inspects its welds from images, welding sound and process data, and corrects them, work the firms say used to depend on highly skilled welders. They say an earlier welding process they developed together cut race-car roll-cage production from two to three weeks to three days.
Achieved whendefect and rework rates for the go-kart frames, before and after, are published or measured by someone outside the two firms.
Claimed by
Yaskawa Electric and GAZOO Racing, October 2026
Open
01
Unseen parts, public tally
An open, filmed test in which a robot welds 30 small steel assemblies it has never seen. It scans each part, finds the joints and writes its own welding program, with no person editing the path. A certified welding inspector then checks every weld, and a sample is cut, polished and etched to show how deep the weld fused. The scans, code, videos and the tally of passes and failures are published.
The briefCredit · unclaimed02
Weld symbols to weld lists
A tool that reads the weld symbols on fabrication drawings, the arrows and marks set out in AWS A2.4 or ISO 2553 that say which joints to weld, with what kind of weld, how big and how long. It writes every weld into a weld list that a shop's weld log or a robot programming tool can import, flagging symbols it cannot read.
The briefCredit · unclaimed
Open 01 · the brief
Unseen parts, public tally
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open, filmed test in which a robot welds 30 small steel assemblies it has never seen. It scans each part, finds the joints and writes its own welding program, with no person editing the path. A certified welding inspector then checks every weld, and a sample is cut, polished and etched to show how deep the weld fused. The scans, code, videos and the tally of passes and failures are published.
Why it is still open
The claims for welding robots that plan their own work come from their makers: Path Robotics quotes first-pass yields of "97%+" and near 99% without a method or a count of new parts, and Yaskawa's self-inspecting robot works on one go-kart frame design. Published research systems, such as the point-cloud planner of Zhou and colleagues (arXiv, 2020), report seam finding and path accuracy, not how many finished welds passed inspection on parts new to the system.
Built with what exists
A cobot with a welding package, such as a Universal Robots UR10e or a Fanuc CRX with a MIG torch, run by a welding school or shop; a depth camera such as an Intel RealSense D405 or a Zivid structured-light camera; Open3D for point clouds and ROS 2 with MoveIt for motion planning; an AI agent that writes and revises readable robot programs and tests them in simulation first (see the Note "Readable robot programs").
First step
Ask a welding school with a cobot to let you scan and weld five T-joint and lap-joint coupons from a depth-camera scan, and log every step a person had to do by hand.
Done looks like
30 assemblies from at least three outside designers, each welded on its first attempt with no human edit to the program, with the scan-to-arc time recorded. Every weld judged by a certified welding inspector against the visual limits of AWS D1.1 or ISO 5817 level C, ten joints cut and etched, and the full tally, videos and code released.
Open 02 · the brief
Weld symbols to weld lists
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A tool that reads the weld symbols on fabrication drawings, the arrows and marks set out in AWS A2.4 or ISO 2553 that say which joints to weld, with what kind of weld, how big and how long. It writes every weld into a weld list that a shop's weld log or a robot programming tool can import, flagging symbols it cannot read.
Why it is still open
Weld takeoffs are still made by people reading the drawing. Commercial drawing readers extract other things: Werk24's API returns dimensions, tolerances, GD&T, materials and title blocks. Two hobby projects on GitHub train YOLO detectors to find weld symbols on drawings, one for 40 AWS symbol types (yolo_welding, 2025) and one for 16 weld types (weldCalloutIdentifier, 2025), but they locate symbols only; they do not read sizes and lengths, build a weld list or publish accuracy against a hand takeoff. Research on reading symbols with deep learning has worked on construction drawings, such as a 2023 comparison of six detectors on bridge and housing drawings by Faltin, Gann and König, and on piping diagrams. Searches in November 2026 found no published tool or public test set for turning weld symbols into a weld list.
Built with what exists
An object detector such as Ultralytics YOLO, starting from the open weld-symbol detectors above; a vision-language model to read the size and length numbers next to each symbol; the symbol rules of AWS A2.4 or ISO 2553 written as code checks; Label Studio for labeling.
First step
Collect 30 public drawings that carry weld symbols, such as open-hardware and public bridge drawings, and make a hand takeoff of every weld as the answer key.
Done looks like
200 drawings from at least five sources, holding 2,000 or more weld symbols, with at least 95% of welds listed correctly (joint, weld type, size, length and side) against a certified welding inspector's hand takeoff, and the test set and scores published with the owners' permission.