Revisiting My 2024 Map of Self-Driving Cars

Revisiting My 2024 Map of Self-Driving Cars

Venture Step E025 tried to map an industry that kept changing names, partners, permits, and ambitions. The episode moved from the DARPA challenges to Waymo, Tesla, Cruise, other development programs, racing experiments, economics, public trust, insurance, and liability. Its strongest question still holds: when software shares or takes over driving, who is responsible for what happens next?

The answer became clearer when I stopped treating "self-driving" as one category.

flowchart LR
    A["Vague self-driving claim"] --> B["Named system and version"]
    B --> C["Driver or ADS role"]
    C --> D["Operational design domain"]
    D --> E["Permit, service, and jurisdiction"]
    E --> F["Safety and incident evidence"]
    F --> G["Case-specific liability questions"]

The DARPA story was about expanding conditions

The episode begins with the 2004 DARPA Grand Challenge, when no vehicle finished the desert route, then moves through later desert success and the 2007 Urban Challenge.

That history is often told as a straight march toward autonomy. A better reading is that each program defined a harder operating environment. Desert navigation, urban traffic rules, intersections, other vehicles, route planning, heat, sensing, communications, and fallback created different problems.

Modern systems still live inside explicit conditions. The term for that boundary is the operational design domain, or ODD. [[What an Operational Design Domain Really Means]] explains why geography is only one field in that record.

Company comparisons needed a smaller unit

E025 compared Waymo, Tesla, Cruise, and several other programs as if companies could be placed on one ladder. That was understandable in 2024, but it mixed driver-assistance products, automated passenger services, tests, partnerships, and abandoned programs.

NHTSA currently distinguishes Level 2 advanced driver assistance from automated driving systems at Levels 3 through 5. Tesla's own current Full Self-Driving (Supervised) page says the features require active supervision and do not make the vehicle autonomous.

Waymo operates a different product and role. Its 2026 company materials describe public fully autonomous passenger service in multiple cities, while California's DMV permit record and CPUC passenger-service record define specific California authorities.

The fair comparison is not "Which company is more autonomous?" It is "What does this system do, who supervises it, where may it operate, under what authority, and what evidence describes that version?"

The industry did not simply converge

The episode was right that programs can disappear, return, or change direction.

General Motors announced in December 2024 that it would stop funding Cruise robotaxi development and focus autonomy work on personal vehicles. GM later acquired full ownership of Cruise and integrated its technical teams.

Zoox took another route. Its current community record says public service is open in Las Vegas and available to selected riders in San Francisco while testing or preparation expands elsewhere. Those are company claims that need matching permits and local authority before being turned into a public status row.

Waymo expanded. The company said in May 2026 that its footprint was growing across 11 cities. That scale matters, but a company announcement is not a regulator's permit and not an independent safety conclusion.

[[Autonomous Vehicle Industry Record]] keeps these states dated instead of declaring a winner.

Safety needs aligned exposure and outcomes

E025 asked whether autonomous systems could be safer than people. That is the right question and an easy one to answer badly.

NHTSA's Standing General Order provides reportable incident data, but the agency warns that the summary counts are not normalized by fleet size, miles, or operational design domain. Reporting entities also differ in how quickly and completely they learn about crashes.

Waymo's Safety Impact hub provides company-authored data, methods, outcomes, locations, and downloadable files. It can support analysis when the system, miles, geography, benchmark, outcome, and uncertainty are aligned. It should not be reduced to a percentage detached from the method.

[[How to Read Autonomous-Vehicle Safety Claims]] shows how to inspect that comparison. [[How to Build an Autonomous-Vehicle Incident Evidence Record]] handles the narrower question of what happened in one event.

Liability starts with the driving role

In the episode, I wondered whether the company or person would be at fault after an automated-driving crash. That framing was too narrow.

At Level 2, the driver remains responsible for supervising and driving. At higher automation levels, the system may perform the entire dynamic driving task within its domain, but liability still depends on facts and law. The vehicle owner, fleet, operator, remote-assistance provider, manufacturer, automation developer, component supplier, maintainer, infrastructure authority, and another road user may all matter.

Insurance also has multiple jobs. It can respond to first-party vehicle damage, bodily injury, property damage, commercial fleet risk, product exposure, cyber events, and recovery against another responsible party, depending on the policy and jurisdiction.

The NAIC autonomous-vehicle topic page describes the evolving insurance and liability frame. The NCSL legislation database tracks state legislation, including insurance and liability topics. Neither decides a real claim.

[[Who Is Liable When Automated Driving Fails]] provides an issue map without assigning fault.

The questions that survived

E025's broad tour is more useful after correction because its underlying concerns survived.

Can the system recognize and stay inside its conditions? What happens at the boundary? How is the human role communicated and monitored? What evidence supports a safety comparison? How are incidents recorded and corrected? Who owns operation, maintenance, updates, and recovery? Which insurer and legal regime apply?

Episode 26 extends the same discipline to AI product claims. E116 applies operational design to shared mobility assets, and E117 examines a digital control system inside physical machinery. Those links should go live only when their destination pages are approved.

The original episode remains on YouTube, Spotify, and its existing Ghost route.

This article was developed with AI assistance from the preserved E025 transcript, current regulator and operator sources, and the linked research record. Dalton Anderson remains the author. It is not legal, insurance, safety, product, or investment advice. Editorial, historical, technical, legal, insurance, source, accessibility, and founder review are required before publication. Publication is not authorized.