Robo-Cab Predictions

[splin]

Robots on the road - how close is our driverless future? Rory Cellan-Jones Technology correspondent

https://www.bbc.co.uk/news/technology-48334449

It was on the motorway near Phoenix, Arizona, that I realised fully driverless cars might be quite a distant dream. And that was because our Google Waymo robo-taxi seemed incapable of leaving that motorway.

...

Myra Blanco, a researcher at the Virginia Tech Transport Institute, in the US, said we would probably see driverless cars in geo-fenced areas in two to five years but she was far more sceptical about full automation.

"That means going from the mountains, rural roadways, all the way to the city - that is going to take a little bit longer, probably potentially a couple of decades away," she said.

Perhaps one of the most intractable problems, as pointed out here several times:

Far from the orderly roads of Arizona, I stood with the transport writer Christian Wolmar at the hectic crossroads outside Holborn Tube station, in central London.

He pointed out that pedestrians would have no hesitation in stepping out in front of driverless cars, knowing they were programmed to stop, and the result would be gridlock.

"Once you set the rule that driverless cars have to effectively kowtow to any pedestrian in the street, and pedestrians begin to learn that, then the whole balance of power in our streets will change," he said.

"The concept just doesn't survive the idea of mixed use streets."

But not to worry - one of the worlds foremost experts in the technology has declared it's imminent:

In the UK, the Chancellor, Philip Hammond, has also promised that "genuine driverless cars" will be on the roads by 2021

 
 
Note he only said "on the road" not "driving on public roads". Although I think what he actually said was "genuine driverless government ..... by 2021"     

[EEVblog]

Every day I drive in traffic I see some situation and I think how will an auto deal with it.

People used to say a chess program would never beat a grand champion, turns out you could do it in 1996 with a disappointingly trivial algoritm as long as you had enough computing power.

Terribly flawed analogy. Chess has small finite set of rules on a fixed playing field. Increased computation and a smart algorithm was always destined to win.

Autonomous cars are playing on an infinitely sized board under an almost infinite set of rules (finite set of fixed road rules, but plus infinite variability when thing change/go wrong)

[magic]

He pointed out that pedestrians would have no hesitation in stepping out in front of driverless cars, knowing they were programmed to stop, and the result would be gridlock.

"Once you set the rule that driverless cars have to effectively kowtow to any pedestrian in the street, and pedestrians begin to learn that, then the whole balance of power in our streets will change," he said.

"The concept just doesn't survive the idea of mixed use streets."

This will be a serious problem and not only with pedestrians. Drivers too violate rules or take risks in order to get somewhere faster and I have heard from Poles living in the EU that we aren't even the worst offenders in this regard
Meanwhile bots will pedantically obey every damn rule about yielding, speed limits etc and generally play it safe to avoid liability and bad PR. That is, until they become completely mainstream and VW starts making them But before, they will be at a constant disadvantage to human drivers exploiting their naivety.
That's why I say it's only going to work in areas where human drivers are already equally docile and there is enough political will to pass and enforce bot-friendly traffic regulations.
Perhaps a mandatory "can't do that Dave" implant in every normal car

[0xdeadbeef]

People used to say a chess program would never beat a grand champion, turns out you could do it in 1996 with a disappointingly trivial algoritm as long as you had enough computing power.

Terribly flawed analogy. Chess has small finite set of rules on a fixed playing field. Increased computation and a smart algorithm was always destined to win.

It's maybe interesting to note that it was once assumed that the problem of a computer playing chess better than a human would be a milestone in creating an artificial intelligence. As chess programs got better and better because of sophisticated algorithms and databases, it became clear that this was a misconception. So it was decided that playing soccer would be a better goal towards AI as it mimics the cooperation of living beings. I'm not sure if everybody has given up on this idea yet, but I think it's pretty clear that you can build a pretty good soccer robot as well without getting one step closer to AI. Actually, in the last decade or so, hardware for neural networks became cheaper and cheaper so it was used commercially even though the roots of neural networks date back to the mid of the 20th century. Now some marketing genius simply declared neural networks to be AI, but it's quite obvious that the extremely simplified simulation of the model (!) of a few brain cells doesn't make it intelligent.
The problem is that many "normal" people fell for this trickery and actually believe that AI has made great progress lately. Indeed most people perceive mathematics, physics, electronics and informatics as some kind of magic anyway and thus can't really evaluate what is possible with the current state of science and what isn't.
Still, not everything human beings do can be easily simulated. And the more complex a scenario is and the less well defined rules exist, the more the simulation actually has to become an artificial intelligence. Nobody would deny that some aspects of driving can be handled by algorithms. Even human drivers are not fully "aware" most of the time. It's the short moments of awareness that are so hard to simulate.

[apis]

Every day I drive in traffic I see some situation and I think how will an auto deal with it.

People used to say a chess program would never beat a grand champion, turns out you could do it in 1996 with a disappointingly trivial algoritm as long as you had enough computing power.

Terribly flawed analogy. Chess has small finite set of rules on a fixed playing field. Increased computation and a smart algorithm was always destined to win.

Autonomous cars are playing on an infinitely sized board under an almost infinite set of rules (finite set of fixed road rules, but plus infinite variability when thing change/go wrong)

Quite right, chess is a so called perfect information game, but it wasn't meant as an analogy but as an example of other things that were deemed impossible for a computer to do by many 20-30 years ago. Since Deep Blue won over Kasparov in 1996 it's been more than 20 years, and computer power have continued to increase exponentially. Things that weren't possible then have become possible now, and that is quite unintuitive to many, and maybe why so many believe self driving cars are impossible.

It's maybe interesting to note that it was once assumed that the problem of a computer playing chess better than a human would be a milestone in creating an artificial intelligence.

Yes, exactly, it was said that you would need full AI to defeat humans in a game like chess (sounds familiar?). Turn's out chess wasn't that hard. Driving is a tougher problem perhaps, but there's nothing magical about that either.

As chess programs got better and better because of sophisticated algorithms and databases, it became clear that this was a misconception. So it was decided that playing soccer would be a better goal towards AI as it mimics the cooperation of living beings.

Hardly. Deep Blue winning over Kasparov was a surprise to most people. IBM's approach was brute force using the traditional algorithms, which came as a disappointment because chess had been so over hyped that people believed that surely you would need some major breakthrough in AI to be able to solve it. Turns out the old algoritms were good enough, they just needed a little bit extra computing power. To solve Go you needed neural nets, so that was a bigger step forward (but it was still old technology that suddenly had enough computing power to make it practically useful).

Robot Soccer is not believed to require full AI. Sure, it's challenging, but it's not the game itself that requires a lot of "thinking", it requires a very agile robot which is a challenging task mechanically. Building a soccer playing robot team requires a wide mix of skills which makes it a nice fun challenge for a team of students.

Nobody would deny that some aspects of driving can be handled by algorithms. Even human drivers are not fully "aware" most of the time. It's the short moments of awareness that are so hard to simulate.

No one is trying to make a full AI to drive cars (unless you listen to Musk). Having a self-aware AI in every car would be crazy and unethical, no one is anywhere near creating full AI. Self driving cars will not be AI, it will be dumb machines that can solve most driving problems autonomously. There will be some driving tasks humans can handle better, but there will be other driving tasks that self driving cars handle better. And it's not about who's best, it's about getting the self driving car "good enough" so that it can be trusted on public roads. (Well, one explicitly stated goal is to reduce traffic accident rates, so they will have to be a lot better in that regard.)

[apis]

Myra Blanco, a researcher at the Virginia Tech Transport Institute, in the US, said we would probably see driverless cars in geo-fenced areas in two to five years but she was far more sceptical about full automation.

"That means going from the mountains, rural roadways, all the way to the city - that is going to take a little bit longer, probably potentially a couple of decades away," she said.

I.e. same as I've been saying and most experts are saying.

[nctnico]

Myra Blanco, a researcher at the Virginia Tech Transport Institute, in the US, said we would probably see driverless cars in geo-fenced areas in two to five years but she was far more sceptical about full automation.

"That means going from the mountains, rural roadways, all the way to the city - that is going to take a little bit longer, probably potentially a couple of decades away," she said.

I.e. same as I've been saying and most experts are saying.

In Sweden they already have a self driving truck on a public road without a driver (there isn't even room for a driver).
https://edition.cnn.com/2019/05/15/tech/einride-self-driving-trucks/index.html

[apis]

In Sweden they already have a self driving truck on a public road without a driver (there isn't even room for a driver).
https://edition.cnn.com/2019/05/15/tech/einride-self-driving-trucks/index.html

Yeah, I've seen that, but it's a very limited trial I believe. It only follows a single pre-programmed route and it's been speed limited (by authorities) to 5 km/h. Still, there might be many truck transport applications where limiting the driving to a few pre-programmed routes would be acceptable.

[Mr. Scram]

With these threads it always seems some people perceive a fundamental difference between artificial and biological neural nets as if biological intelligence isn't the result of using the proper hardware and optimising it to the hilt. There often also seems to be the assumption that something has to be roughly equivalent to human intelligence to be called intelligent.

[0xdeadbeef]

Yes, exactly, it was said that you would need full AI to defeat humans in a game like chess (sounds familiar?). Turn's out chess wasn't that hard. Driving is a tougher problem perhaps, but there's nothing magical about that either.

Driving in a complex real world environment (i.e. one that a human driver could handle) is a totally different task than mastering a game with fixed rules.

Hardly.

Actually, since the mid 90s, quite a lot of people uttered something like "Robocup is the next AI challenge" .
http://theconversation.com/why-football-not-chess-is-the-true-final-frontier-for-robotic-artificial-intelligence-62296
https://medium.com/syncedreview/having-notched-impressive-victories-over-human-professionals-in-go-atari-games-and-most-recently-30b88ee363e9

To solve Go you needed neural nets, so that was a bigger step forward (but it was still old technology that suddenly had enough computing power to make it practically useful).

Hard to say if Go couldn't have been beaten with traditional algorithms just because nobody managed (up to now). Anyway, it's still a very well defined environment with very strict rules and without realtime requirements (turn based like chess), so something like playing soccer on the level of humans is actually much more challenging.

Hardly.
Robot Soccer is not believed to require full AI. Sure, it's challenging, but it's not the game itself that requires a lot of "thinking", it requires a very agile robot which is a challenging task mechanically. Building a soccer playing robot team requires a wide mix of skills which makes it a nice fun challenge for a team of students.

Of course the first step is using tiny little robots with small fields and simplified rules. However, the realtime requirements and the nearly endless possibilities in every single moment make this a challenge much closer to autonomous driving than chess or Go. Playing soccer on the level of human players is of course a mechanical/robotic problem as well and by simplifying the robots, playing environment and rules, the problem might be diminished to something than can be solved by an algorithm. However, the final goal of Robocup is to develop a team of independent robots than can beat the human world champion team.
Actually, probably nobody can tell if this can still be done by algorithms but I kinda doubt it. Still, a soccer game is of course a very well defined challenge in a confined space with clear rules and therefore much less complex than autonomous driving in a real world traffic scenario (rush hour in major city during a hailstorm with roadworks and what not).

No one is trying to make a full AI to drive cars (unless you listen to Musk).

Well, an artificial intelligence on the level of a human driver is needed to cope with every possible situation in a realworld environment as good as a human driver. IMHO, whoever denies this, underestimates the complexity.

Having a self-aware AI in every car would be crazy and unethical, no one is anywhere near creating full AI.

Self-awareness is a totally different issue. Of course, at this point, I guess nobody on this planet could answer if achieving true AI is even possible without self awareness or not. Besides, there is no black and white about self-awareness. Also animals have different levels of self-awareness. Then again, it's actually pretty hard to prove self-awareness. You could easily program a robot or program to pretend self-awareness but it would be incredibly hard to prove if a black box AI is fully self aware or just pretends to be.

Self driving cars will not be AI, it will be dumb machines that can solve most driving problems autonomously.

Nah, this will never work reliably in the real world. It will work though in a restricted environment but that just doesn't really help for true autonomous driving.

There will be some driving tasks humans can handle better, but there will be other driving tasks that self driving cars handle better. And it's not about who's best, it's about getting the self driving car "good enough" so that it can be trusted on public roads. (Well, one explicitly stated goal is to reduce traffic accident rates, so they will have to be a lot better in that regard.)

A dumb machine will never be good enough to be entrusted with people'e lives on public roads.

[EEVblog]

Every day I drive in traffic I see some situation and I think how will an auto deal with it.

People used to say a chess program would never beat a grand champion, turns out you could do it in 1996 with a disappointingly trivial algoritm as long as you had enough computing power.

Terribly flawed analogy. Chess has small finite set of rules on a fixed playing field. Increased computation and a smart algorithm was always destined to win.
Autonomous cars are playing on an infinitely sized board under an almost infinite set of rules (finite set of fixed road rules, but plus infinite variability when thing change/go wrong)

Quite right, chess is a so called perfect information game, but it wasn't meant as an analogy but as an example of other things that were deemed impossible for a computer to do by many 20-30 years ago. Since Deep Blue won over Kasparov in 1996 it's been more than 20 years, and computer power have continued to increase exponentially. Things that weren't possible then have become possible now, and that is quite unintuitive to many, and maybe why so many believe self driving cars are impossible.

No one is saying they are "impossible".
But they are currently flawed and will be for a very long time, the infinite variability in conditions practically guarantee this.
Your statement about chess most certainly is an analogy, otherwise it's pointless to mention.
Many (most?) people make the huge mistake that because self driving cars can do albeit impressive things now, that perfection is only a few years away. But it doesn't work like that with open ended complex problems in an infinitely variable playing field like self driving cars will have endure. Smarter people do understand this which is why there is much talk about them only truly being practical when they have their own dedicated lanes, or once they are the majority on the road etc.

It's the same thing with humanoid robots. People see the Boston Dynamics robots etc and think that perfection is only a few years away. Not even close.

[EEVblog]

There will be some driving tasks humans can handle better, but there will be other driving tasks that self driving cars handle better. And it's not about who's best, it's about getting the self driving car "good enough" so that it can be trusted on public roads. (Well, one explicitly stated goal is to reduce traffic accident rates, so they will have to be a lot better in that regard.)

A dumb machine will never be good enough to be entrusted with people'e lives on public roads.

Perception will be very important in the adoption of autonomous cars.
People expect a much higher degree of perfection from computers, and when people start dying in autonomous cars (as they have been already) public opinion will not be kind.
And the argument that "but self driving cars are statistically safer" will not hold water, people won't care, they'll still trust humans more for a very long time.

[0xdeadbeef]

People expect a much higher degree of perfection from computers, and when people start dying in autonomous cars (as they have been already) public opinion will not be kind.
And the argument that "but self driving cars are statistically safer" will not hold water, people won't care, they'll still trust humans more for a very long time.

And after the first few deaths, hell will rain down.

[Mr. Scram]

And after the first few deaths, hell will rain down.

Although I suspect you're right that's actually quite ridiculous. Human drivers are fairly terrible at driving. Our sensor array is average at best and easily swamped by environmental factors. Our reaction time is atrocious and the worst part is probably that we aren't properly aware of our own capabilities and limitations. Worldwide millions of people die in traffic each year and in the US alone 40000 people meet their untimely end each year. We seem to accept that so why would automated systems be held to unrealistic standards? Likely because of the same reasons we suck at driving in the first place. The public at large isn't very good at understanding statistics and our own limitations. We'll believe we could do better even if it kills us.

[apis]

No one is saying they are "impossible".
But they are currently flawed and will be for a very long time, the infinite variability in conditions practically guarantee this.

They will never be perfect, that is the nature of the unpredictable and infinitely variable conditions in real life (and neither are humans). Even the Waymo CEO has said that "autonomy always will have some constraints". And you will always be able to point out situations that autonomous cars doesn't handle as well as humans (because we work very differently). But they don't have to be as good as humans at all tasks in order to be safe and useful.

Your statement about chess most certainly is an analogy, otherwise it's pointless to mention.

I wasn't saying chess and driving was analogous, but that the notion that it was impossible for a machine to beat humans in chess is analogous to the notion that it is impossible for a machine to drive safely on public roads.

Many (most?) people make the huge mistake that because self driving cars can do albeit impressive things now, that perfection is only a few years away. But it doesn't work like that with open ended complex problems in an infinitely variable playing field like self driving cars will have endure.

I think the huge mistake is assuming it has to be perfect. It only has to be cost-effective and "good enough". Human drivers are far from perfect either.

Smarter people do understand this which is why there is much talk about them only truly being practical when they have their own dedicated lanes, or once they are the majority on the road etc.

Having dedicated lanes would make things a lot easier for sure. The people who have been involved in developing self driving cars are pretty smart, I wouldn't be so quick to dismiss what they are doing.

It's the same thing with humanoid robots. People see the Boston Dynamics robots etc and think that perfection is only a few years away. Not even close.

The Boston Dynamics robots are impressive from a mechanical point of view, but they still have a long way to go before you can compare such a robot to a human body. We have more than 200 bones, 360 joints, 650 linear actuators, massive sensor arrays, and a lot of parallel computing power that's been optimised over millions of years. It's going to take a lot of time and money to beat that. (Which is why robot soccer is hard, not because the game strategy is difficult.)

[apis]

Actually, since the mid 90s, quite a lot of people uttered something like "Robocup is the next AI challenge" .
http://theconversation.com/why-football-not-chess-is-the-true-final-frontier-for-robotic-artificial-intelligence-62296
https://medium.com/syncedreview/having-notched-impressive-victories-over-human-professionals-in-go-atari-games-and-most-recently-30b88ee363e9

The strategy part of soccer isn't hard, it's creating a mechanical robot that can match a human that is hard. As I wrote in the previous post we have more than 200 bones, 360 joints, 650 linear actuators, massive sensor arrays, and a lot of parallel computing power that's been optimised over millions of years. It's going to take a lot of time and money to beat that.

Or in the words of the first article you linked to:
"The tasks involved in playing football, although much more intuitive to humans than chess or Go, are a major challenge for robots. Technical problems of hitherto unimaginable complexity have to be solved: timing a kick while running, identifying the ball against a glaring sun, running on wet grass, providing the robot with sufficient energy for 45 minutes’ play, even the materials that go into constructing a robot can’t disintegrate during a forceful game. Other problems to be solved will define important aspects of our life with robots in the future: when a robot collides with a human player, who can take how much damage? If humans commit fouls, may a robot foul back?"

With autonomous driving the mechanical part isn't a problem, because both humans and computers drive the same kind of machine (i.e. cars) which we already know how to make.

To solve Go you needed neural nets, so that was a bigger step forward (but it was still old technology that suddenly had enough computing power to make it practically useful).

Hard to say if Go couldn't have been beaten with traditional algorithms just because nobody managed (up to now). Anyway, it's still a very well defined environment with very strict rules and without realtime requirements (turn based like chess), so something like playing soccer on the level of humans is actually much more challenging.

Soccer strategy really isn't hard. One reason go is harder than chess is because the search space is much much larger than chess, so just adding a little bit more computing power wouldn't have been effective any time soon. Of course, they still used some traditional techniques in Alpha Go, but it was really the neural nets that was the key.

Actually, probably nobody can tell if this can still be done by algorithms but I kinda doubt it.

Let's take a look at how computers perform at imperfect information games that is much more difficult than soccer from a strategy point of view:
EDIT: Sorry, I was intending to post this newer follow up video (I'll leave the old one below for those interested)

Tihis is an earlier video from 2018:

No one is trying to make a full AI to drive cars (unless you listen to Musk).

Well, an artificial intelligence on the level of a human driver is needed to cope with every possible situation in a realworld environment as good as a human driver. IMHO, whoever denies this, underestimates the complexity.

A self driving car doesn't need "to cope with every possible situation in a realworld environment as good as a human driver".

Self driving cars will not be AI, it will be dumb machines that can solve most driving problems autonomously.

Nah, this will never work reliably in the real world. It will work though in a restricted environment but that just doesn't really help for true autonomous driving.

"True" autonomous driving is a red herring, what does it even mean? Humans also drive in a restricted environment; we only drive along certain roads, we try to follow well defined rules and we sometimes (quite often) fail miserably and need help from others to get out of a tricky situation.

There will be some driving tasks humans can handle better, but there will be other driving tasks that self driving cars handle better. And it's not about who's best, it's about getting the self driving car "good enough" so that it can be trusted on public roads. (Well, one explicitly stated goal is to reduce traffic accident rates, so they will have to be a lot better in that regard.)

A dumb machine will never be good enough to be entrusted with people'e lives on public roads.

Except, they already are. People are already testing without human drivers on public roads.

[Mr. Scram]

Could we perhaps limit the point for point quotes and replies? This doesn't really make the thread legible or enjoyable.

[apis]

Our sensor array is average at best and easily swamped by environmental factors. Our reaction time is atrocious and the worst part is probably that we aren't properly aware of our own capabilities and limitations. Worldwide millions of people die in traffic each year and in the US alone 40000 people meet their untimely end each year.

That is definitely true compared to a self driving car, we are limited to using our eyes and ears (and maybe a little bit of feedback from the steering wheel and acceleration/vibration). A computer can use lidar, radar and can use cameras that see in the thermal regions of the spectrum, and have no dead angles. That (and the fact that computers doesn't get distracted/drunk/sleepy/etc) is why I think it is a safe to assume that self driving cars will be able to reduce traffic accidents significantly.

[0xdeadbeef]

The strategy part of soccer isn't hard, ...

Who talked about strategy? It's about the player judging the situation and making decisions in realtime.

Or in the words of the first article you linked to:
"Technical problems of hitherto unimaginable complexity have to be solved: timing a kick while running, identifying the ball against a glaring sun, r [...]"
With autonomous driving the mechanical part isn't a problem, because both humans and computers drive the same kind of machine (i.e. cars) which we already know how to make.

Even in the cherry picked quote, you left have these statements about timing and object recognition. Additionally, a humanoid soccer robot playing against a human (since this is the goal!) would have to predict human behavior - exactly as an autonomous car. It's just that the complexity is much higher for the car as there are so many possibilities for humans and animals and what not to interact with the car.

Soccer strategy really isn't hard. One reason go is harder than chess is because the search space is much much larger than chess, so just adding a little bit more computing power wouldn't have been effective any time soon.

You continue to ignore that even a large number of turns with fixed movement rules is a total piece of cake compared to any real world movement scenario.

Let's take a look at how computers perform at imperfect information games that is much more difficult than soccer from a strategy point of view:

Again, that's a virtual environment where nothing unexpected can happen. Letting aside if the computer performs image analysis at all and that the movement control is very basic: also the actions of each unit are clearly defined etc. Agreed, it's real time, but very, very far away from the complexity of the real world.

A self driving car doesn't need "to cope with every possible situation in a realworld environment as good as a human driver".

Sure it has to. Else the whole thing is totally pointless and must be restricted to confined areas where the possible interactions with the real world are limited to a degree the dumb car can handle.

"True" autonomous driving is a red herring, what does it even mean? Humans also drive in a restricted environment; we only drive along certain roads, we try to follow well defined rules and we sometimes (quite often) fail miserably and need help from others to get out of a tricky situation.

Humans can follow rules but also decide when they better shouldn't. If something blocks the lane, a human driver will be able to react in a sensible way. A dumb car will just stop and block the traffic. Now even this can cause deaths if the car is not able to let an ambulance pass.

There will be some driving tasks humans can handle better, but there will be other driving tasks that self driving cars handle better. And it's not about who's best, it's about getting the self driving car "good enough" so that it can be trusted on public roads. (Well, one explicitly stated goal is to reduce traffic accident rates, so they will have to be a lot better in that regard.)

The few things that a computer can handle better like brake assist can be added without replacing the human driver. An autonomous car that replaced the human driver has to be at least as good as a human driver in every aspect.

Except, they already are. People are already testing without human drivers on public roads.

AFAIK that's not even true. I'm not aware of any tests on public roads without a human driver who can take over at any time. And if this driver is not 100% focused, this can result in a dead pedestrian as we've learned.

[apis]

The strategy part of soccer isn't hard, ...

Who talked about strategy? It's about the player judging the situation and making decisions in realtime.

Sounds like a definition of real time strategy to me.

Or in the words of the first article you linked to:
"Technical problems of hitherto unimaginable complexity have to be solved: timing a kick while running, identifying the ball against a glaring sun, r [...]"
With autonomous driving the mechanical part isn't a problem, because both humans and computers drive the same kind of machine (i.e. cars) which we already know how to make.

Even in the cherry picked quote, you left have these statements about timing and object recognition. Additionally, a humanoid soccer robot playing against a human (since this is the goal!) would have to predict human behavior - exactly as an autonomous car. It's just that the complexity is much higher for the car as there are so many possibilities for humans and animals and what not to interact with the car.

It wasn't cherry picked, it was the relevant bit showing that the difficult part isn't in analysing the game and planning the strategy, it's in building the robot. Chess and Go is much more difficult games than soccer in terms of strategy.

Since when was timing harder for a computer than a human? 

Computer vision is hard which is why self driving cars have lidar so that they don't have to rely on computer vision algorithms for the safety-critical parts. For less critical tasks like predicting human (and other animal) behaviour they use neural nets. They don't have to be 100% accurate in that case and neither are humans so it works well.

You continue to ignore that even a large number of turns with fixed movement rules is a total piece of cake compared to any real world movement scenario.

Because going from discreet to continuous isn't what's challenging.

Let's take a look at how computers perform at imperfect information games that is much more difficult than soccer from a strategy point of view:

Again, that's a virtual environment where nothing unexpected can happen. Letting aside if the computer performs image analysis at all and that the movement control is very basic: also the actions of each unit are clearly defined etc. Agreed, it's real time, but very, very far away from the complexity of the real world.

Now you're moving the goal posts (pun intended). The difficult part of robot soccer isn't the strategy part, it's creating a robot with similar speed, agility and endurance as a human pro football player (i.e. incredibly difficult). Identifying the ball and other players (computer vision) wouldn't be hard for a robot today either, running and kicking around a ball accurately is. But the mechanics is not a problem for a self driving cars, since we already have the mechanical parts (the cars).

A self driving car doesn't need "to cope with every possible situation in a realworld environment as good as a human driver".

Sure it has to. Else the whole thing is totally pointless and must be restricted to confined areas where the possible interactions with the real world are limited to a degree the dumb car can handle.

No it really doesn't, that is your mistake. It only have to be convenient and cost efficient enough. We human drivers are also restricted to confined areas where the interactions with the real world are limited to a degree the dumb humans can handle.

"True" autonomous driving is a red herring, what does it even mean? Humans also drive in a restricted environment; we only drive along certain roads, we try to follow well defined rules and we sometimes (quite often) fail miserably and need help from others to get out of a tricky situation.

Humans can follow rules but also decide when they better shouldn't. If something blocks the lane, a human driver will be able to react in a sensible way. A dumb car will just stop and block the traffic. Now even this can cause deaths if the car is not able to let an ambulance pass.

Humans can't handle all situations either (not even sensibly). It's not black and white. It's acceptable if self driving cars need human assistance in a small number of situations.

Except, they already are. People are already testing without human drivers on public roads.

AFAIK that's not even true. I'm not aware of any tests on public roads without a human driver who can take over at any time. And if this driver is not 100% focused, this can result in a dead pedestrian as we've learned.

It's actually old news by now.

Oct 30, 2018:
"We’re excited to announce that the California Department of Motor Vehicles (DMV) has just granted Waymo the first permit in the state to begin driverless testing on public roads.

This permit is the result of new DMV regulations that took effect in April, which allow companies to apply for fully driverless testing within carefully defined limits, and is the product of nearly ten years of testing in California by Waymo’s team. It’s the first time that California has allowed tests on public roads of fully driverless cars ― that is, without a test driver sitting in the driver’s seat.

Familiar ground
Fully driverless testing is the latest step in the path Waymo has been on since 2009, when we first began working on self-driving technology at Google. Since then we’ve driven over ten million autonomous miles on public roads across 25 cities. California will join our driverless testing program that’s already been happening in Phoenix, Arizona since last year.
"
https://medium.com/waymo/a-green-light-for-waymos-driverless-testing-in-california-a87ec336d657

Video from 2017 (Arizona testing begins)


Video from 2018 (California testing begins)

[nctnico]

Humans can follow rules but also decide when they better shouldn't. If something blocks the lane, a human driver will be able to react in a sensible way. A dumb car will just stop and block the traffic. Now even this can cause deaths if the car is not able to let an ambulance pass.

I don't think you have a driver's license or (while being in a car) pay any attention to the stupid stuff other people do while driving a car (or attempting to do so). Especially the Saturdays and Sundays are bad.

[0xdeadbeef]

I don't think you have a driver's license or (while being in a car) pay any attention to the stupid stuff other people do while driving a car (or attempting to do so). Especially the Saturdays and Sundays are bad.

Well, maybe you should train that thinking thing a bit then. Anyway, of course there a dumb drivers as there are lots of dumb people. Still, even a stupid driver has a strong interest in getting forward instead of just waiting forever. Just watch what happens if a traffic light gets stuck in the red phase. Human drivers will tolerate this for a few minutes and then the first ones will either ignore the red light or turn around or maybe even call the police. Nobody will just stand there and wait forever.

[Mr. Scram]

Well, maybe you should train that thinking thing a bit then. Anyway, of course there a dumb drivers as there are lots of dumb people. Still, even a stupid driver has a strong interest in getting forward instead of just waiting forever. Just watch what happens if a traffic light gets stuck in the red phase. Human drivers will tolerate this for a few minutes and then the first ones will either ignore the red light or turn around or maybe even call the police. Nobody will just stand there and wait forever.

We're all fairly familiar with watchdog timers. You could even imagine a situation where a car calls for help when stuck in a situation it can't see a solution for. You don't need to go fully flawlessly automated in one go.

[apis]

That's why everyone is planning on ride-share or taxi services I believe. There will always be an operations center that is in contact with all the cars. If there is a problem the computer or the passengers can call for help. Operators can then remote control or guide the car until it can manage by itself again, or in worst case send a replacement car and a technician.

[Mr. Scram]

That's why everyone is planning on ride-share or taxi services I believe. There will always be an operations center that is in contact with all the cars. If there is a problem the computer or the passengers can call for help. Operators can then remote control or guide the car until it can manage by itself again, or in worst case send a replacement car and a technician.

It's not unlikely the car can call for help itself and instructions can be given to it indirectly, with the car figuring out the exact execution. If it's commanded to run a red light the car should be able to figure out what a good moment to do so is. This is pretty much how the more recent Mars rovers are driven, minus the traffic lights of course.