A car that stays in its lane, maintains distance and stops for obstacles is not a hard engineering problem in 2026. Systems that do it competently have shipped for years. The difficulty was never the ordinary case.
What the last part consists of
Situations that are individually rare and collectively constant. A traffic officer waving cars through a red light. A mattress in the fast lane. A cyclist making eye contact and going anyway. Roadworks with contradictory signage. A child’s ball entering the road before the child does.
Each of these is unusual. Any given hour of driving contains at least one, and a system that fails on them is not a system anyone should deploy.
Why more data has diminishing returns
Training on driving data means each additional mile mostly contains situations the system has already seen. The valuable examples – the genuinely strange ones – become proportionally rarer as the dataset grows.
This is why fleets accumulating hundreds of thousands of unsupervised miles is a real milestone and not a solution. The miles establish that the ordinary case holds. They say much less about the tail.
The negotiation problem
Driving is partly social. A merge in dense traffic requires signalling intent and reading whether it has been accepted. Human drivers do this continuously and unconsciously. An autonomous vehicle that follows the rules exactly is predictable and often stuck, because the rules assume a cooperation the rules cannot describe.
Where the honest progress is
Constrained deployment. Specific cities, mapped in detail, in weather within known limits, at hours with known traffic patterns. That is how every credible service currently operates, and it works.
The gap between that and a vehicle that handles any road anywhere is not one more model generation. It is a different problem.
What that means
Autonomous transport arrives as a service in specific places long before it arrives as a feature in a car you own. The services expanding now are real. The consumer version being months away has been months away for a decade.