[sofixa]

This reminds me of Waymo's approach to self-driving cars. Paraphrasing, but basically they found that progressively adding self-driving to help human drivers is bad, because it leads to the humans becoming complacent and not paying enough attention. Therefore they decided on an all or nothing approach, where their cars would be only and entirely self-driven.

[typewithrhythm]

This always seemed like a bit of bull from waymo. It's not an easy problem to work with existing manufacturers to give a better and or cheaper solution... Especially when there are established competitors with efficient verification and validation processes (that every manufacturer requires).

They decided it wasn't worth explaining that their techniques don't generalise to a driver assist. It would not be good or cheap enough to be worth developing the compliance and integration frameworks.

[jowday]

The first thing Waymo tried building (way back when, circa 2010 or so) was highway based driver assist in the style of Autopilot. They did a ton of testing with it and didn’t like how quickly their testers stopped paying attention despite a ton of instruction not to. I’ve seen clips from these tests.

It’s also possible they shifted direction because the long term vision of robotaxis is much more lucrative.

[typewithrhythm]

The thing is, you can tune driver assist to whatever form you like; but then you have a true comparison of function with established competitors.

I just don't believe the approaches for high autonomy (especially at the time) actually could make a cost effective assist system.

And for whatever reason they decided to push the we didn't like the driver behaviour message, rather than actually talking about what was actually plausible to achieve in the driver assist space.

[warble]

Maybe I'm being ignorant about something here but isn't paying less attention the whole point?

[sofixa]

Unless you assume that self-driving software is perfect, no, it really isn't. That's the whole problem - the drivers would get complacent, so when there's an issue, they'd be caught by surprise and wouldn't be able to react.

[mook]

Isn't the point to pay _no_ attention? The difference is when an accident occurs, was the person in the car at fault for not vigilantly watching everything.

[renewiltord]

You want the curve of total attention to be always above a baseline human in an unassisted car. The car can do some attention and the human can do some. But if the sum of the two falls below the threshold, you’re in trouble.

[Veserv]

You can already do that by just closing your eyes and letting Jesus take the wheel. No, the point is doing so while maintaining safety.

It is materially less safe to operate a ADAS while distracted than driving manually. Humans are exceptionally good drivers on average, only encountering minor crashes on timeframes measured in years to decades. As such, if safety critical ADAS errors occur more frequently than every ~100,000 miles and you are attentive in less than 100% of all such occurrences, you are operating your vehicle multiple times more dangerously than the average driver (which is a number that includes drunks and distracted drivers).

That is why it is critical to deliberately downplay the capabilities, to avoid wishful over-reliance, and enforce strict driver awareness (through techniques such as driver monitoring) to avoid operating multi-ton killing machines in ways that are multiple times more dangerous to the occupants, other drivers, and pedestrians. Without that, people are prone to over-generalization of safety capabilities, extrapolating that a single success means robust, continued success thousands to tens of thousands of times in a row.

[simondotau]

That statement from Waymo always struck me as deeply uninsightful because it was really just a more complicated way of saying “self driving systems need to be good enough to drive the car on their own” which isn’t just obvious, but tautological.