[lisper]

An even more extreme example: suppose someone on the sidewalk suddenly whips out a bazooka and shoots it at you. Does your failure to anticipate this contingency count as a failure?

"Failure" must be defined with respect to a particular model. If you're driving in the United States, you're probably not worried about bazookas, and being hit by one is not a failure, it's just shit happening, which it sometimes does. (By way of contrast, if you're driving in Kabul then you may very well be concerned with bazookas.) Whether or not you want to worry about drunk pedestrians and avoid them at all possible costs is a design decision. But if you really want to, you can (at the possible cost of having to drive very, very slowly).

But no reasonable person could deny that avoiding collisions with stationary obstacles is a requirement for any reasonable autonomous driving system.

[andosa]

Way to dodge the question. And how did we get from always knowing when you're failed to "just drive very, very slow when", when dealing with situations that human drivers deal with all the time.

Let's not pretend that anticipating potentially dangerous behaviour from subtle clues is some once-in-a-lifetime corner case. People do this all the time when driving -- be it a drunk guy on the sidewalk, a small kid a tad bit too unstable when riding a bike by the roadside, kids playng catch nex to the road and not paying attention, etc etc. Understanding these situation is crucial in self driving if we want to beat the 1 fatality per 100M mile that we have with human drivers. For such scenarios, please explain how the AI can always know when it failed to anticipate a problem that a normal human driver can.

[lisper]

> how did we get from always knowing when you're failed to "just drive very, very slow when", when dealing with situations that human drivers deal with all the time

You raised this scenario:

> there's a very drunk person on the sidewalk. As a human driver, you know he might act unexpectedly so you slow down...

I was just responding to that.

> Let's not pretend that anticipating potentially dangerous behaviour from subtle clues is some once-in-a-lifetime corner case.

I never said it was. All I said was that "failure must be defined with respect to some model." If you really want to anticipate every contingency then you have to take into account some very unlikely possibilities, like bazookas or (to choose a slightly more plausible example) having someone hiding behind the parked car that you are driving past and jumping out just at the wrong moment.

The kind of "failure" that I'm talking about is not a failure to anticipate all possible contingencies, but a failure to act correctly given your design goals and the information you have at your disposal. Hitting someone who jumps out at you from behind a parked car, or failing to avoid a bazooka attack, may or may not be a failure depending on your design criteria. But the situation in the OP video was not a corner case. Steering into a static barrier at freeway speeds is just very clearly not the right answer under any reasonable design criteria for an autonomous vehicle.

My claim is simply that given a set of design criteria, you cannot in general build a system that never fails according to those criteria, but you can build a system that, if it fails, knows that it has failed. I further claim that this is useful because you can then put a layer on top of this failure-detection mechanism that can recover from point failures, and so increase the overall system reliability. If you really want to know the details, go read the thesis or the paper.

These are not particularly deep or revolutionary claims. If you think they are, then you haven't understood them. These are really just codifications of some engineering common-sense. Back in 1991, applying this common sense to autonomous robots was new. In 2018 it should be standard practice, but apparently it's not.