Is there a standard that these cars have to meet?
Really there should be an obstacle course of some kind based on the most common anomalies leading to accidents. Partially obscured road signs, worn paint markings, obscured sight lines caused by dense parking, driving on a snowed over surface, etc.
Such a standard for motor vehicles would result in the loss of licensing for a large majority of Americans. Most people would fight against such testing standards as a condition for driving on public roads. It might not be a bad idea. I just don’t think it’s feasible.
There's no legislated standard in the US, nor pre-sale verification that a vehicle has met most of the other automotive standards. That's just not how the US regulatory environment works.
There's the beginnings of a standard for L2/L3 vehicles in Europe, and ongoing discussions about what L4 standards could look like, but neither addresses "adversarial" situations.
China is going for what's sometimes called "L3.5" - hands off wheel, eyes off road, driver may have to take over if alerted, but if the driver doesn't, the vehicle must safely stop and pull over if possible.
The trouble with "driver assistance" systems is the handoff to the driver. That's not reliable. Results from the aviation community indicate that it takes about ten seconds after an autopilot disconnect before the pilot has stable control of the aircraft. "L3.5" accepts that. The < 2 second response required by some L2 systems does not.
Real world data results are so much more important than an obstacle course to overfit for.
While I would like to see that just for the data, from a regulatory standpoint all we really need are the manufacturers being liable in a crash if their cars were driving. The insurance savings of their cars being safer will create the incentive they need.
If it’s anything like my San Francisco DMV driving test, they did a loop around Fell and Oak St (both one way streets with no opposing traffic).
I passed because I didn’t kill anyone in the span of eight blocks.