The Self-Driving Vehicles Thread

[Horse]

If - big 'if' - the system is usually better than a human and can alert the human in sufficient time when required, how would you feel about that?

Not a big 'if', a f---ing HUGE 'if', and one most unlikely to be fixed ever.

If it could recognise the problem that far ahead it could just stop, so almost by definition the 'problem' scenarios where it can't deal with stuff will be rather urgent. The reality is that cars will hand back control when a crash s inevitable, so just shifting the blame really.

The reality is - probably - that the first implementation will be in environments where there is a simple choice: go or slow/stop.

The sort of situations I presume that you are imagining (pile-ups on a motorway where the AV is thundering along in lane 3, please give a few examples if I'm way off the mark) are not going to be happening for some considerable time, if at all. It's - with my very limited understanding of how development is going - more likely (and this is a guess) - will have had a "computer says no" moment long before, handing back control when a certain level of complication and risk has been reached and not exceeded.

My - as I said, limited - understanding is that systems won't be designed that say "Ding Dong. Crash imminent. Your turn." Trolley bus, again.

[wheelnut]

My feeling, based on nothing more than gut, is that manufacturers will abandon efforts to develop fully autonomous cars. I cba to dig out the link but the head of Citroen has already said they won’t happen in our lifetimes.

Yes, we’ll have limited environment exceptions but the fact is that they will not scale.

It appears humans aren’t as bad as driving as we thought.

[Horse]

My feeling, based on nothing more than gut, is that manufacturers will abandon efforts to develop fully autonomous cars.

It appears humans aren’t as bad as driving as we thought.

I posted a quote from a major developer a few days ago saying roughly that. Pretty sure I was saying the same back on TRC too, and that last/first mile implementation will be the first things you'll see.

Are humans good? Or is it more that the developers underestimated the complexity? Actually there's a third element: cost. Until very recently, lidar systems were horribly expensive. Also, the extra processing from hi-def cameras etc means hefty computing is required

[Cousin Jack]

It appears humans aren’t as bad as driving as we thought.

Humans (and their driving) are incredibly complex systems. Complexity demands even more complexity to manage it successfully.

[Horse]

It's probably 'reasonable' to say that there might have been a slight underestimation by ADS developers of the complexity of driving

However, what the systems are beginning to achieve is impressive. And, like any robot designed for a task, they will usually do it consistently. The problem is identifying 'too complex' early enough. I wouldn't be too surprised if the scenarios that you might imagine as too complex are not the ones concerning developers.

[Cousin Jack]

I agree completely about closed and controlled environment stuff, autonomous vehicles work well and are probably far more consistent than fallible humans.

My concern is much more about fallible humans placing far too much trust in semi-autonomous vehicles and, despite warnings, relying on them far beyond their safe environment.

As for complexity, any hint of human control adds massive complexity, so any environment with multiple humans will be a nightmare for full automation. Especially if those humans are not constrained within steel cages!

[Horse]

My concern is much more about fallible humans placing far too much trust in semi-autonomous vehicles and, despite warnings, relying on them far beyond their safe environment.

As for complexity, any hint of human control adds massive complexity, so any environment with multiple humans will be a nightmare for full automation.

IIRC some car manufacturers have said that they may bridge L4, ie include L3 features in vehicles that they sell but have L5 as their next step.

[wheelnut]

IIRC some car manufacturers have said that they may bridge L4, ie include L3 features in vehicles that they sell but have L5 as their next step.

I think there could be some tinkering with existing driver aids but I can’t see mainstream adoption of any technology which fundamentally changes the way we drive any time soon.

[wheelnut]

Some Darwin Award nominees

https://www.theguardian.com/technology/ ... eat-police

[Horse]

Some Darwin Award nominees

https://www.theguardian.com/technology/ ... eat-police

Yes, just been reading about that. Have to wonder whether they had survived previous 'stunts' but, this time, the car enabled their demise.in fact, perhaps the AI is set to identify such people and 'delete' them (hopefully not any bystanders though)?

[The Spin Doctor]

I agree completely about closed and controlled environment stuff, autonomous vehicles work well and are probably far more consistent than fallible humans.

My concern is much more about fallible humans placing far too much trust in semi-autonomous vehicles and, despite warnings, relying on them far beyond their safe environment.

As for complexity, any hint of human control adds massive complexity, so any environment with multiple humans will be a nightmare for full automation. Especially if those humans are not constrained within steel cages!

I'm with CJ on this.

Designers placing too much trust in fallible humans? Partly that, partly failing to understand that humans simply cannot stay on task (as a teacher would put it) when there's no stimulation to stay on task.

I think the entire 'semi-autonomous' idea is simply not a runner in the hugely complex driving environment. The point that I've been making all along is that you can't 'hand back' control to drivers and expect them to react instantaneously. It's simply not going to happen for all the reasons of cognitive neuroscience that I'm sure we're aware of.

The system needs a) to bring a driver's conscious decision-making mind into focus, b) give him / her time to recognise what happening, c) decide what needs to be done, then finally d) do it.

That's how brains work, bottom line.

Another idea I've seen floated is have massive penalties for failing to be alert behind the wheel when a car is in self-driving mode. I'm sure we also know that 'after the fact' deterrents for failing to do what we're supposed to do are spectacularly ineffective.

[Horse]

The point that I've been making all along is that you can't 'hand back' control to drivers and expect them to react instantaneously.

So what if the driver isn't expected to take back control instantaneously? What if the system was designed to operate on its own for several - 'many' even - seconds. What it won't commit itself to a situation that it can't cope with?

[The Spin Doctor]

So what if the driver isn't expected to take back control instantaneously? What if the system was designed to operate on its own for several - 'many' even - seconds. What it won't commit itself to a situation that it can't cope with?

And round and round we go - circular argument from you again.

As I've said dozens of times on this and similar threads, the uncontrolled environment frequently doesn't give 'the system' time to operate on its own for several seconds. Any busy town street, any crowded motorway at speed.

The question you should be asking is "how will you design a system that KNOWS when it CAN commit to a situation", when the driving environment is so incredibly fluid.

[Horse]

So what if the driver isn't expected to take back control instantaneously? What if the system was designed to operate on its own for several - 'many' even - seconds. What it won't commit itself to a situation that it can't cope with?

And round and round we go - circular argument from you again.

Perhaps answer the question instead of blaming me?

You seem to assume that you can identify and predict problems that either the machine can't, or if it can, then it can't deal with them.

How about picking a few real world examples (and, as I have said, several times, I am not involved in developing these things, I am sharing what I know when I can).

So what about 'pedestrian runs out from in front of a bus'.

Could you predict that, would it be a surprise, might you be startled so not react promptly?

I know that exact scenario has been used. The intent is to take AVs beyond 'simply' steering between the obstacles. So, potentially, the AV would not only identify a blind area where a surprise might occur but, potentially, be ready to react and, If that pedestrian did run out, it's unlikely that the AV would be struck rigid by 'startle'.

It could assess the situation, approach accordingly, react appropriately. No need to pass back control.

I've posted this before.



If you don't want to watch the whole video, FF to about 15 minutes.

[Horse]

Another idea I've seen floated is have massive penalties for failing to be alert behind the wheel when a car is in self-driving mode.

There you are - you know more about this than I do.

Although I thought we already had similar laws, however effective they are.

[The Spin Doctor]

And round and round we go - circular argument from you again.

Perhaps answer the question instead of blaming me?

I'm saying I've already answered it... many times.

You seem to assume that you can identify and predict problems that either the machine can't, or if it can, then it can't deal with them.

Nope. I haven't said that. What I HAVE said - repeatedly - is that "you can't have a machine that appears to be in full control suddenly throwing its mechanical hands in the air" and expect a driver to take control.

If a machine can deal with ALL driving situations, fine. If it can deal with LIMITED driving situations without recourse to the driver, fine.

But as I have said - oh so many times - the halfway house where a driver is supposed to be alert and ready to take control at a moment's notice - as in the Tesla crash above - is a nonsense and I don't think the technological challenges will ever be overcome.

Ask me again in five, ten, and (hopefully) twenty years.

[Horse]

What I HAVE said - repeatedly - is that "you can't have a machine that appears to be in full control suddenly throwing its mechanical hands in the air" and expect a driver to take control.

Why do you assume that's when it will happen? Presumably you have a reason for that belief?

[Horse]

And, as I have probably asked before, why do you think that none of this has occurred to anyone involved in developing these contraptions?

[The Spin Doctor]

And, as I have probably asked before, why do you think that none of this has occurred to anyone involved in developing these contraptions?

I didn't say that either.

What I have said is that I doubt their ability to solve that puzzle.

[The Spin Doctor]

What I HAVE said - repeatedly - is that "you can't have a machine that appears to be in full control suddenly throwing its mechanical hands in the air" and expect a driver to take control.

Why do you assume that's when it will happen? Presumably you have a reason for that belief?

Tesla.