The MachineRecord
Incident register for commercial robotics

Every incident
Every consequence

We keep a case per event, not a record per article. A collision, the investigation that follows, the recall, the lawsuit and the settlement stay in one file. A copy of each source is kept, and linked beside it with the date it was taken, because half of these links will not open in two years.

Cases
280
Sources
619
Escalated
39
Updated
2026-09-16
Cases by year of incident

The curve bends up

025507510020162017201820192020202120222023202428202576202688
cases on record reached regulator, recall, lawsuit or settlement2026 is a partial year. Search feeds also surface recent publications more readily than old ones, so part of the rise is coverage, not incidence.
Escalation index

Not who is dangerous. Whose incidents reach consequences.

Fleet sizes are not public, so we cannot report incidents per thousand machines and will not pretend otherwise. What we can measure is internally normalised: the share of a company's recorded incidents that reached a regulator, a recall or a court.

CompanyCasesEscalatedShareSince
Zoox95
56%
2024
Cruise103
30%
2022
Tesla152
13%
2019
Starship Technologies91
11%
2019
Waymo695
7%
2023
Unitree Robotics110
0%
2025

Companies with fewer than 8 cases are left out of the index: at that sample size a single case moves the share by double digits. Currently outside it: Avride 4, Boston Dynamics 4, Intuitive Surgical 4, iRobot 3, Ecovacs 3, Baidu 3, Coco Robotics 3, Serve Robotics 3, and 33 more.

Registry

Latest cases

The open web yields a handful of genuinely new events per week. Most additions are developments on cases already on file, or older events entering the register for the first time.

Filed with the regulator

What the operators report about themselves

Every company running driving automation on American roads must file a report with NHTSA for each crash. Those filings are public, and they are a spreadsheet of 116 columns, which is why almost nobody reads them. The register holds all of them.

Reports on file
1437
Hospitalisation or worse
51
Operators
17
Cities
68

By place: San Francisco 378, Los Angeles 247, Austin 189, Phoenix 109, Atlanta 73, Dallas 66

Methodology

What this data does not say

A register is useful exactly as far as it is honest about its own limits. Coverage is English-language, so incidents in Shenzhen or Osaka appear only when someone wrote about them in English. Minor faults never reach the news at all. And a denominator exists for only a handful of cases, where a recall disclosed how many units were out there.

Read the methodology
Escalation chain

Every case carries a position on this chain. The red line on a case, a company card or a table row is the same measurement.

  1. First report
  2. Followed up
  3. Investigation opened
  4. Regulator acts
  5. Recall issued
  6. Fine issued
  7. Claim filed
  8. Criminal charge
  9. Paid or resolved

Everything above is free and always will be.

Every case, every timeline and every source link stays open and indexable. Paid products cover convenience: company dossiers as a finished PDF, alerts on the companies you track, and export with an API.