Hacker News Reader

Comment by gdulli | Hacker News Reader

> The system should have started applying brake at this point. If a 3500lb vehicle can't decide what it is about to impact, it needs to slow down (to a stop if necessary).

I suspect there's a big problem in the other direction, too. If the system starts to brake every time it thinks it might need to, it will happen all the time. It might see false positives like this all the time and that's why it doesn't act on them right away.

If it's (appropriately) conservative about starting to brake it will be braking/slowing all the time. If it's not conservative, people or cars will occasionally get hit. The former could make people uncomfortable or carsick or create some subtler danger by stopping short needlessly. The latter might mostly work and mostly not kill people, except for when it does.

View on HN

Then your system is not ready to be on the road. We're taking about metal box weighting at least a ton, launched at dozen of miles an hour on roads where they can hit people.

"They had to not react quickly because they keep mis-detecting things" is really not an acceptable justification. If that's the why (and it probably is) that car should never have been on the street.

Exactly my point. I'm not apologizing for the software. I have severe doubt that software is anywhere close to making safe and appropriate decisions in the context of real world driving with open ended conditions about the makeup of the road, full range of obstacles and edge case conditions, weather, lighting, etc.

The scariest thing is that it's so fucking familiar, this idea of technologists solving a range of simple cases and assuming the solution extrapolates to success in the real world. We cannot get into these cars if that kind of software development feels anything like everyday software development, where random web site bugs can be tolerated.

No, the scariest thing is that capital may eventually push through a launch of these crappy devices before they're ready. And they'll never be ready.

Why do you think they'll never be ready?

Because I don't think computer vision will ever be good enough to be safe for this purpose.

We trust human vision for this purpose; is there anything human vision is doing that would be impossible to emulate with a computer, given enough research and development?

What exactly do you mean by "enough?" "Enough to solve the problem" is a tautology.

You're the one who used the word "enough" in "ever be good enough to solve the problem." What did you mean by it? Are you claiming that human perception cannot ever be synthetically matched?

I imagine most self-driving systems will have a higher success rate than human drivers. Uber's may not of been one of those systems though.

If self-driving cars were remotely possible, it would be obvious, because Google Maps would be insanely perfect. That's the bottom line as far as I'm concerned. Every time it directs me to do something stupid, I think, what if it were telling my car directly?

I'm sure the situation will improve eventually, but I'm guessing we're still several decades away and that it'll take advances in both sensor technology and software before driverless cars work at the same level as a human driver in all conditions. In perfect weather, on perfect roads, with minimal traffic and zero surprises the tech works mostly OK now, but those are basically lab conditions and not the reality most of us drive in today.

How do you feel about Tesla releasing their fully self driving software soon? They seem to have a very low accident rate so far.

I would want better reliability than "it shouldn't be a problem...until it is." That's pretty much the same reason my 82yo mom used and let me be the first to say you don't want to be driving next to her car (she doesn't drive anymore).

As a preliminary step to any regulatory approval Tesla should release every byte of data from their tests so we can analyze the scenarios and events that the software has dealt with, so we can second-guess them. I seem to remember a common criticism of Tesla is that it's kinda shitty to work there and I don't think the best work comes out of an environment like that.

We should know for sure what they/you mean by "seem to" and "very low." Trade secret protection is insignificant when public safety is involved.

Are 2 errors better than 35,000 human errors?

Depends on the errors!

Along the same lines, if they can figure out how to drive a car, they can figure out how to give instructions to human drivers that are far, far better than the state of the art. So, when they release their revolutionary navigation system, I'll try it. Not buying a car from them though.

If we apply this to a human, let's say I'm getting old and my eyes aren't as good any more, I don't just keep driving like nothing changed, I need to find a way to see better, because the risk of getting into an accident is higher. If the car's system cannot tell what an object is, you don't just assume it's going to be ok, you need to either get better sensors or find some other real solution.

>* suspect there's a big problem in the other direction, too. If the system starts to brake every time it thinks it might need to, it will happen all the time. It might see false positives like this all the time and that's why it doesn't act on them right away.*

Oh no, it doesn't work and Travis & friends have been lying for years. Wake me when there are consequences for the follies of rich white guys.

>If the system starts to brake every time it thinks it might need to, it will happen all the time. It might see false positives like this all the time and that's why it doesn't act on them right away.

if your car always sees empty space ahead as non-empty space, you probably shouldn't let the car on the road until you fix that. Once you fixed that, if the car sees that the space ahead is non-empty, even it can't classify it, it should slow down well before and warn the driver and continue with the braking if the driver is asleep/watching youtube. It is AZ, there is no rain nor snowflakes falling which would mislead the lidar. An object ahead - slow down and stop if you can't navigate it.

Presenting it as the AI-hard issue of misclassification is just an attention misdirection from and a whitewhashing/laundering of the foundational issue of knowingly letting car on the road with missing basic safety level of "don't hit objects in front of the car". Similar to the Boeing blaming 737 MAX crashes on the failed sensor.

>It is AZ, there is no rain nor snowflakes falling

This is so understated. This car was basically driving in ideal conditions at night for a self-driving car. If it can't avoid a hit under these conditions, it clearly wasn't ready to be on the road.

> If the system starts to brake every time it thinks it might need to, it will happen all the time.

One thing that self driving will "give back" is time to the occupant.

And the thing that most human driven accidents are correlated to damage; speed.

Reply on news.ycombinator.com