Industries / Transport and vehicles / Automotive
Final assembly
Final assembly fits the painted body with its powertrain, wiring, interior, glass, wheels and fluids while it moves down a line, and the stations along the line are fed the right parts in the right order.
5
Milestones achieved, from 1913 to 2018.
3
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1913
Ford's moving assembly line
At Highland Park, Michigan, Ford moved the car past workers at fixed stations instead of sending workers around a parked car. The labor in a Model T fell from about 12½ man-hours to about 1½, by The Henry Ford museum's account.
LessonThe faster line had a human cost; by late 1913 Ford's labor turnover was 380 percent.
Credit
Henry Ford and Ford Motor Company's engineers
1955
Workers can stop the line
Toyota linked stop buttons to andons, lit boards that show which station has a problem, on its Crown assembly line at the Honsha plant. A worker who finds a fault stops the line, and work restarts once the cause is fixed. Toyota calls the principle jidoka, "automation with a human touch".
LessonStopping to fix a cause costs less than passing a defect down the line.
Credit
Toyota, which credits Taiichi Ohno with building the Toyota Production System
1984
NUMMI: same workers, new system
In February 1984 Toyota and General Motors formed NUMMI, a joint venture in GM's closed plant at Fremont, California. John Shook, who worked for Toyota on the venture, wrote that absenteeism fell from over 20% to a steady 2% and that quality went from GM's worst to GM's best within a year.
LessonIn Shook's words, "the only thing that changed was the production and management system".
Credit
Toyota and General Motors
1985
Hamtramck's automation backfires
GM opened its Detroit-Hamtramck plant in 1985 full of new automation. Lines stopped for hours while software was fixed, and computers sent body parts to the wrong cars; output improved after the plant removed technology that did not work.
Lesson"Consistency of manufacture must come before automation", as the Tuck case study puts it.
Credit
General Motors; a Dartmouth Tuck School case study ties the push to chief executive Roger Smith
2018
Tesla scales back automation
Tesla missed its target of building 2,500 Model 3 cars a week, reaching 2,020, and removed a complex conveyor system from the line. Elon Musk wrote on 13 April 2018 that "excessive automation at Tesla was a mistake".
LessonThe Hamtramck lesson had to be learned again 33 years later.
Credit
Tesla, as reported by Fortune
Watch for
Humanoid kitting at Mercedes-Benz
Kitting means picking parts from bulk bins into trays in the order each car on the line needs them. Google DeepMind says its Gemini Robotics 2.5 model helped Apptronik's Apollo 2 humanoid robot pick and pack car parts on active lines at Kecskemét, Hungary, and Sindelfingen, Germany, with a 96% success rate; the post gives no attempt count and does not define a success (see the Note "A real production line").
Achieved whenMercedes-Benz or an outside party reports the robots on routine shifts, or publishes a measured success rate with attempt counts.
Claimed by
Google DeepMind, November 2026
Figure 03 sequencing at BMW
Sequencing means putting parts on a cart in the order cars will reach the line. BMW says Figure 03 humanoids will take unsorted parts from large containers and sort them into a sequencing trolley at its Spartanburg, South Carolina plant, and gives no success rate or cycle time. BMW says the earlier Figure 02 loaded sheet-metal parts for welding in the body shop and supported production of more than 30,000 BMW X3s over ten months.
Achieved whenBMW reports routine use of Figure 03 with output figures, as it did for Figure 02.
Claimed by
BMW Group and Figure AI, June 2026
Atlas sequencing at Hyundai Georgia
Boston Dynamics opened a training center inside Hyundai Motor Group Metaplant America near Savannah on 21 September 2026, where its Atlas humanoid learns parts logistics and sequencing. Hyundai plans to deploy Atlas at the plant from 2028, "subject to continued technology validation", and Boston Dynamics says applications will extend to component assembly by 2030.
Achieved whenAtlas sequences parts in routine production, confirmed by someone outside the two companies.
Claimed by
Hyundai Motor Group, July 2026, and Boston Dynamics, September 2026
Open
01
Robot fits a door harness
A bench rig that holds one real front-door inner panel, trim removed, with its own wiring harness, both from a scrapyard. The door is unpowered, and doors that carry airbag parts are skipped. A robot arm takes the harness off a hook, lays it along its path, pushes each plastic clip into its hole and plugs each connector until it latches.
The briefCredit · unclaimed02
Public tally for robot kitting
A small station that copies the Mercedes-Benz and BMW pilot jobs: bins of three real part types (brackets, clips and fasteners from a dealer's parts counter) and trays with slots for three car versions. A robot fills trays from an order list. The definition of success is written before the run, and every attempt is filmed and logged.
The briefCredit · unclaimed
Open 01 · the brief
Robot fits a door harness
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A bench rig that holds one real front-door inner panel, trim removed, with its own wiring harness, both from a scrapyard. The door is unpowered, and doors that carry airbag parts are skipped. A robot arm takes the harness off a hook, lays it along its path, pushes each plastic clip into its hole and plugs each connector until it latches.
Why it is still open
Gartner analyst Pedro Pacheco calls cabin assembly and wire harness installation "the final frontier" of assembly automation, and consultant Werner Geiger puts automation of Western European final assembly at 20 to 35 percent (January 2026). A 2025 study by KIT, ArtiMinds and others automated only the plugging of connectors on an Audi center console, reaching 88% to 98% success per connector, and states that "wire harness installation remains a largely manual process". ARENA2036, a research campus in Stuttgart, is building a demonstrator for this with BMW, Mercedes-Benz and others and lists no results.
Built with what exists
A collaborative robot arm such as the Universal Robots UR5e, which the KIT and ArtiMinds study used; or, for a cheaper first attempt, an SO-101 arm with LeRobot (see the Note "No download"); a camera; NIST's free build files for its Task Board #4, a wire-harness practice board; AI agents that write robot programs a person can read (see the Note "Readable robot programs").
First step
Buy one door with its harness, photograph and number every clip hole and connector, and film a person fitting it five times.
Done looks like
Twenty filmed, uncut attempts on the same door, with a public table recording for each attempt the clips seated (each pulled by hand afterwards), connectors latched (pull test), any pinched wire, minutes taken and human touches, plus a person doing the same job five times for comparison.
Open 02 · the brief
Public tally for robot kitting
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A small station that copies the Mercedes-Benz and BMW pilot jobs: bins of three real part types (brackets, clips and fasteners from a dealer's parts counter) and trays with slots for three car versions. A robot fills trays from an order list. The definition of success is written before the run, and every attempt is filmed and logged.
Why it is still open
The leading claims for this job publish no attempt-level data. DeepMind gives 96% without attempt counts or a definition of success, BMW and Figure give no success rate for Figure 03, and Boston Dynamics gives no performance figures for Atlas.
Built with what exists
An SO-101 arm with LeRobot (see the Note "No download"); an open robot model such as Runway Praxis-1 (see the Note "Open robot model weights") or NVIDIA GR00T 2 (see the Note "Five demonstrations per task"); phone cameras; a public spreadsheet.
First step
Write the success definition and the order list, and buy the three part types.
Done looks like
200 filmed tray fills across the three versions, with a public log that marks each as pass, wrong part, wrong slot, dropped or missed grasp, with the time per tray, and a person doing the same 200 as a baseline.