From what I've seen I wouldn't trust autonomous cars to "understand" all situations. I would trust Waymo cars to understand enough to avoid hitting anything (at a level better than a human), at the risk of being rear-ended more often. Everything I've seen from Tesla and Uber has given me significantly less confidence than that.
The argument I've always heard is that an autonomous systems will outperform humans mostly by being more vigilant (not getting distracted, sleepy, etc.) rather than using detectors with superhuman reaction times. Obviously, whether or not this outweighs the frequency of situations where the autonomous system gets confused when a human would not is an empirical question that will change as the tech improves.
And, like, once the systems are mature and hardware has evolved and so forth, I think that's right. It will. It might today under certain circumstances, like if it gets a really unambiguous sensor input (or it might not).
But I've never heard anyone actually associated with a driverless car program assert that their vehicles have superhuman reaction times today, and the vehicles drive extremely cautiously. I think it's likely that due to their difficulty in understanding their sensor readings, if you look at total time necessary to make a course correction from the point when an obstruction first could be noticed by an alert human driver, driverless cars are not winning and may be substantially losing in at least some cases.
I agree with the sentiment though. This has been a major selling point for this technology, but it has not been sufficiently demonstrated yet.
(It certainly didn't predict the crash "seconds before", as the car accelerated into the imminent colission, until <1sec before collision.
Something that is endemic in HN discussions of driverless cars is commenters who dramatically overestimate the dangers of humans driving. Like an order of magnitude or more. You see it all over this thread, in which people imagine that killing one pedestrian in like let's say 10 to 20 million miles driven (with those miles overwhelmingly done in unchallenging conditions) constitutes "vastly safer than human drivers" rather than "vastly less safe than human drivers."
Toyota has enough UX data that they added "brake assist", turns out that a lot of accidents happen when the user stomps the brake, but then releases, just never presses it hard enough to come to a complete stop, or was trained before anti lock brakes.
https://www.youtube.com/watch?v=grcuorbrYxA
Another scary thing about braking is that most cars probably don't have their brake fluid changed often enough; most owners and even dealers treat it as a lifetime interval, when it may be closer to an annual thing interval.
Driverless cars simply don't have the mileage yet to prove they are safer than driverful cars. But there's also no indication they are less safe, based on casualty rate so far.
There is an indication that they're less safe! They have (very conservatively) 5x the number fatalities per mile driven! Now, look, the error bars on that estimate are of course massive. It is plausible that they are much safer and they just rolled the dice and got unlucky. But this is data, as long as you include the error bars.