Industries / Transport and vehicles / Automotive
Quality
Quality is the work of checking parts and finished cars against their specifications, running the shared rulebooks that car makers and suppliers follow, and recalling cars when a defect reaches customers.
6
Milestones achieved, from 1950 to 2024.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1950
Deming teaches Japanese managers
The Union of Japanese Scientists and Engineers (JUSE) invited W. Edwards Deming, an American statistician, to Japan in July 1950, where he taught an eight-day course on statistical quality control. He gave his royalties to JUSE, which used them to found the Deming Prize.
Credit
W. Edwards Deming; JUSE
1994
One rulebook for suppliers
Chrysler, Ford and General Motors replaced their separate supplier rules with a shared standard, QS-9000, first published in August 1994. It later became ISO/TS 16949, which IATF 16949:2016 replaced. The International Automotive Task Force (IATF), the car makers' group that owns the standard, added Chery as a member in June 2026.
LessonOne shared rulebook spared suppliers from meeting many separate customer requirements.
moderncasting.comiatfglobaloversight.orgiatfglobaloversight.org
Credit
Chrysler, Ford and GM; the IATF
2000
TREAD Act after Firestone
After Firestone tires failed, mostly on Ford Explorers, the US TREAD Act, signed on 1 November 2000, required makers to report foreign recalls within five days and set up an early-warning system for defect data. The National Highway Traffic Safety Administration (NHTSA) later noted that it had received about 46 complaints about the tires in ten years, while Firestone held 193 injury claims, 2,288 property-damage claims and 66 lawsuits that it did not have to report.
LessonThe warning signs sat in company files the regulator could not see.
Credit
US Congress; NHTSA
2014
GM fined for slow reporting
General Motors agreed in May 2014 to pay NHTSA a $35 million fine for failing to report an ignition-switch defect in time. An internal review led by former US attorney Anton Valukas found what chief executive Mary Barra called "a pattern of incompetence and neglect".
LessonUnder US law, slow reporting of a defect is a violation in its own right.
Credit
NHTSA; Anton Valukas
2015
Takata: largest US recall
In May 2015 Takata agreed to expand its US recall of airbag inflators to nearly 34 million vehicles, and in May 2016 regulators added 35 to 40 million more inflators. The propellant, ammonium nitrate, breaks down with time, moisture and heat.
LessonA part can pass every factory check and still fail years later from age and climate.
Credit
NHTSA and the US Department of Transportation; Takata
2024
AI cameras in every plant
AIQX is BMW's in-house system that reads camera images on the line and flags missing parts, wrong variants and misplaced labels, so workers can fix them before the car moves on. Axis Communications, the camera supplier, reports that it is deployed in all BMW Group plants worldwide.
Credit
BMW Group
Watch for
Glue check steers the robot
Coherix says its 3D GlassMaster Tracker inspects the bead of urethane glue that holds a windshield in place as a robot lays it, and adjusts the robot's path in real time; its chief executive says it removed up to five hours of weekly line downtime for several customers. No car maker or plant is named.
Achieved whenA named car maker or an outside auditor reports downtime or scrap figures from a named plant.
Claimed by
Coherix, September 2026
Complaints warn years early
Researchers from DrivePulse, a company, and Georgetown University say their statistical method flagged warning signs in NHTSA owner complaints for 19 of 24 vehicles later recalled for the GM ignition-switch defect (79%), a mean of 7.43 years before the recalls. This is the authors' own test on one case, and they write that broader validation remains an open challenge.
Achieved whenAnother group reproduces the method on other recalls and publishes its catch rate and false alarms.
Claimed by
Tommy Jones and colleagues, at NHTSA's ESV conference in Toronto, May 2026
Open
01
Recall early-warning benchmark
A public test for methods that flag defects from owner complaints. Take NHTSA's public complaint file, freeze copies as they stood before each of a set of past recalls, and add matched vehicle and component pairs that never led to a recall. Score any method on how many recalls it flags, how early, and how many false alarms it raises. This uses old data only and makes no safety decisions.
The briefCredit · unclaimed02
Hear a connector latch
A public set of sound recordings of real car wiring connectors being pushed together, some fully latched and some deliberately left half-seated, mixed with factory noise at set levels, plus a detector that tells the two apart. All connectors are unpowered.
The briefCredit · unclaimed
Open 01 · the brief
Recall early-warning benchmark
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test for methods that flag defects from owner complaints. Take NHTSA's public complaint file, freeze copies as they stood before each of a set of past recalls, and add matched vehicle and component pairs that never led to a recall. Score any method on how many recalls it flags, how early, and how many false alarms it raises. This uses old data only and makes no safety decisions.
Why it is still open
The DrivePulse and Georgetown authors write that their method "lacks ground truth labels for most events" and that "many recalls occur without complaint spikes", and they describe earlier studies as "typically evaluated offline" on "limited subsets of the vehicle market". No shared benchmark turned up in this research.
Built with what exists
NHTSA's public complaint and recall files; Python; a small language model run locally to read complaint text (see the Notes "One GPU to fine-tune" and "A first stable number for local AI"); a second model that flags doubtful labels for a person to check (see the Note "Seven more flagged").
First step
Download NHTSA's complaint and recall files, pick ten recalls from 2015 to 2020, and build their frozen snapshots by filtering complaints by the date NHTSA received them.
Done looks like
At least 50 past recalls and 50 non-recall controls, each with a frozen complaint snapshot; published scoring code; and a results table for at least two methods, one a simple complaint-count baseline, giving recalls caught, median months of warning and false alarms per recall caught.
Open 02 · the brief
Hear a connector latch
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of sound recordings of real car wiring connectors being pushed together, some fully latched and some deliberately left half-seated, mixed with factory noise at set levels, plus a detector that tells the two apart. All connectors are unpowered.
Why it is still open
A TE Connectivity patent granted in January 2019 says a connector "can be partially mated during a car assembly process" and still pass conventional electrical tests, then come loose from vibration in service, and that a worker may not hear the latch click over factory noise. Renault's public AutoVI set of assembly-line images (2024) has no connector category, and this research found no public recording set or published miss rate.
Built with what exists
Connector pairs cut from salvaged harnesses; a USB microphone array; Hitachi's public MIMII recordings of industrial machine sounds for background noise; a small open audio classifier such as Google's YAMNet, fine-tuned; a pull gauge to confirm each latch.
First step
Buy five connector pairs from a salvage yard, record 50 full and 50 half matings of one pair on a phone, and check whether the click shows in a spectrogram.
Done looks like
At least 1,000 recorded matings across at least five connector types, about half of them half-seated and each confirmed by a pull test, under an open license, and a published table of the share of half-seated matings the detector misses and the share of good ones it rejects, at three noise levels.