Robo-Cab Predictions

[nctnico]

Why would a self driving car need to park? In many cities driving around on the slowest route will be cheaper compared to paying for the parking space.

Besides that many people are really dumb when it comes to finding a parking space. On a lower or upper floor you can usually park right next to the elevator and get out of the parking garage quicker. Instead many people like to wait until someone is leaving a space (causing a traffic jam and delay) because they are to neurotic to look further. A self driving car can take these metrics into account an find the most optimal parking space. In the future self driving cars are likely able to talk to the 'empty space' detection system and drive to the nearest available parking space.

[Brumby]

I would argue that real autonomous driving (in a real world traffic situation you could encounter today in a car) could only work reliably with real artificial intelligence (and not some brainless neuronal network giving somewhat nondeterministic results) so it could react to unexpected situations at least as good as a human driver.
So even ten years from now sounds pretty much optimistic unless there is some unexpected breakthrough in AI.

Yep, in all the hype this gets forgotten, they aren't even close to what a human driver can do in terms of the unexpected. Won't even happen in 10 years, I'm calling that one too.

I think you are wrong here

You are entitled to your opinion.  So is Dave.

and it is time you back these kind of claims with some factual information.

Why?  Dave has expressed his opinion based on his own criteria.  He may be right or he may be wildly off the mark.  There's no need for him to offer any "factual information" to support his opinion, especially since it has been offered as his own personal speculation.

At least I don't think so.

[nctnico]

Well this is an engineering forum so it would be nice to have a discussion which is based on some factual information and not wild guesses.

[Brumby]

I don't think the EV market will go away like that.  Certainly, it will lose a high profile player, but there are other manufacturers who are in the game to some degree - and with some bold legislative targets in place in some areas and the general tendency away from the love affair with petroleum fuels, it will continue.  There will be an impact, but it won't be terminal.

Tesla isn't the biggest player in the EV market today. Not by a long shot.

I never said nor intimated that Tesla was anything more than high profile.  Being high profile means that your brand is known to a great many people.  It has nothing directly to do with production volume, capitalisation or share price.

[Brumby]

Well this is an engineering forum so it would be nice to have a discussion which is based on some factual information and not wild guesses.

So - you would put a straight jacket on free thinking?

Look back and see how good engineers have been in applying their formal thinking to futurist pursuits.  The writers of Star Trek have done a better job in some respects.  (But that's just my opinion.)

[Brumby]

Why would a self driving car need to park? In many cities driving around on the slowest route will be cheaper compared to paying for the parking space.

Let's see...
 * Wasted energy
 * Extra wear and tear on the vehicles
 * Increased maintenance costs
 * Unnecessary traffic increasing congestion and slowing down other vehicles actually going somewhere
 * Increased presence on the roads means an increased risk of collision

... that's just from off the top of my head.

[EEVblog]

I would argue that real autonomous driving (in a real world traffic situation you could encounter today in a car) could only work reliably with real artificial intelligence (and not some brainless neuronal network giving somewhat nondeterministic results) so it could react to unexpected situations at least as good as a human driver.
So even ten years from now sounds pretty much optimistic unless there is some unexpected breakthrough in AI.

Yep, in all the hype this gets forgotten, they aren't even close to what a human driver can do in terms of the unexpected. Won't even happen in 10 years, I'm calling that one too.

I think you are wrong here and it is time you back these kind of claims with some factual information.

Just my engineering opinion based on what I see, experience, and can easily imagine in terms of possible complex scenarios.
I've seen the tracking data on the best autonomous cars and they miss plenty of basic stuff. Even the best systems in pretty simple driving scenarios average something like only a few hours before they require human intervention. It's not hard to imagine countless scenarios that would confuse even the best systems and how these systems aren't even close to learning how to deal with them.
And people are predicting the roads will be filled with fully autonomous cars within 1-2 years. Bullshit.

[wilfred]

Why would a self driving car need to park? In many cities driving around on the slowest route will be cheaper compared to paying for the parking space.

Besides that many people are really dumb when it comes to finding a parking space. On a lower or upper floor you can usually park right next to the elevator and get out of the parking garage quicker. Instead many people like to wait until someone is leaving a space (causing a traffic jam and delay) because they are to neurotic to look further. A self driving car can take these metrics into account an find the most optimal parking space. In the future self driving cars are likely able to talk to the 'empty space' detection system and drive to the nearest available parking space.

I've observed that parking lot behavior too. But my point was not just in a regulated parking garage but also in the street and wherever a less clearly delineated cluster of parking spaces exist.
As for an auto knowing where a free empty space is, well I'd like to see it argue with a human driver who thinks they saw it first. It will possibly literally demonstrate "punching their lights out".

My basic argument is that where humans and autos interact there are situations that are not part of the programmed intelligence of the machine. As long as that remains so then autos are simply robots that need an artificially constrained environment to operate in. They will not be ready for the real world soon. And I am betting the cost will be too high and the world will move past them being a viable option. Other requirements of burgeoning urban development will drive us away from the current car lifestyle. I already see younger people for whatever reason opting to remain without a car for longer than was done when I was their age. Continually spending billions to build expensive and controversial tunnels and other road infrastructure to relieve local congestion bottlenecks can only go on so long. You cannot spend your way out of road congestion and although autos were meant to make roads flow more smoothly that also is a temporary solution. Cities would sprawl a lot less if they didn't need to accommodate so much roads. Cities like villages  before cars can only grow to a size that requires a tolerable time spent commuting. It's not distance it's time spent commuting.  That time defines the size of the city because it defines how much roads and cars are needed and can fit into the available space. I reckon we'll hit the limits before we develop autos and when we hit the limits no amount of automation will help.

[apis]

Even the best systems in pretty simple driving scenarios average something like only a few hours before they require human intervention.

Then you are misinformed. Anyone testing self driving cars in California are required to publish their disengagement rate by law. Waymo currently has the best disengagement rate (for 2018):
"Across the millions of urban miles we’ve driven on California roads, our disengagement rate dropped to 0.09 per 1,000 self-driven miles in 2018 (or 1 disengage per 11,017 miles self-driven)."
GM was second with about double that. You can see a complete table here:
https://www.therobotreport.com/waymo-autonomous-vehicles-apple/
(Note that Tesla isn't even listed because apparently they aren't even testing self driving cars in California.)

And people are predicting the roads will be filled with fully autonomous cars within 1-2 years. Bullshit.

Only Elon Musk says that, and yes that is bullshit. But Waymo will probably begin testing without safety drivers in Phoenix within the next 5 years. Naturally it will be many more years before the roads are filled with autonomous cars.

[apis]

Even if they can drive well and if they can park in a space, how will they go finding a space to park?

What the technology leaders are planing are taxi services, a robo-cab won't need to find a parking space in a crowded parking lot, they just have to pick up or let of their passengers before they drive off again. They will have to park at times, but they can then find a spot in a less crowded area. The same is actually true for a privately owned autonomous car, it could drop you off at your destination and then drive somewhere where there are free parking spots. (But I don't think we will see privately owned and operated autonomous cars for a long time.)

Is this even part of the programming? I know I use lots of clues to find a spot including looking to see if someone is in the drivers seat and may be about to put their seat belt on. I also look at people with a shopping bag and stalk them in case they are leaving.

While the cars don't rely on neural networks to determine if there is an obstacle on the road in front of the car (that is what they use the lidar for), they do use neural network for a lot of other tasks. Determining if someone is in the drivers seat and are about to leave is a perfect task for a neural network, and Waymo have claimed they can do exactly that. They are very likely not 99% accurate, but neither would a human be, and they don't have to. Neural networks really excel at such tasks, often demonstrating better than human performance.

The cars have to be able to recognise parking signs to ensure they don't park and get a ticket. Who gets the ticket?

Not sure why recognising parking signs would be an issue? They also have detailed maps that could list parking spots. The ticket goes to the owner/operator of the vehicle.

[Mr. Scram]

Weed smoking Musk has promised exactly that, putting them on the streets next year. Yeah, another visionary thing. With cannabis the term "visionary" takes a new meaning.

That guy is full of it. Practical self landing rockets and attractive electric cars? Get real.

[EEVblog]

I've observed that parking lot behavior too. But my point was not just in a regulated parking garage but also in the street and wherever a less clearly delineated cluster of parking spaces exist.
As for an auto knowing where a free empty space is, well I'd like to see it argue with a human driver who thinks they saw it first. It will possibly literally demonstrate "punching their lights out".
My basic argument is that where humans and autos interact there are situations that are not part of the programmed intelligence of the machine. As long as that remains so then autos are simply robots that need an artificially constrained environment to operate in. They will not be ready for the real world soon.

THIS.

All we ever see is these autonomous cars doing (relatively) basic stuff, as advanced and impressive as that is of course, again it's not anything close to what a human is capable of.
Can't say I've seen one get out of the way of an ambulance. Pull over when the cops point their finger. Notice that someone is walking toward their car and is about to hop in and so wait for a minute.
And I could go on and on and on for countless scenarios.

And I am betting the cost will be too high and the world will move past them being a viable option.

It will be interesting to see in what markets it's a fad, and what markets is finds success in.

[EEVblog]

While the cars don't rely on neural networks to determine if there is an obstacle on the road in front of the car (that is what they use the lidar for)

The Tesla doesn't have a LIDAR, and Musk recently said any company who uses one is doomed.

[Mr. Scram]

Just my engineering opinion based on what I see, experience, and can easily imagine in terms of possible complex scenarios.
I've seen the tracking data on the best autonomous cars and they miss plenty of basic stuff. Even the best systems in pretty simple driving scenarios average something like only a few hours before they require human intervention. It's not hard to imagine countless scenarios that would confuse even the best systems and how these systems aren't even close to learning how to deal with them.
And people are predicting the roads will be filled with fully autonomous cars within 1-2 years. Bullshit.

Can you share your sources? What I'm able to find paints quite a different picture of about 0,09 interventions per 1000 miles, which translates to about one intervention every 18000 kilometres. I do agree with your initial assessments, if only because the law will prevent a rapid adoption. Regulations are sluggish. Hybrid and partial solutions are likely to be seen first, with things like trucks on highways making the transistion first.

https://www.engadget.com/2019/02/13/waymo-self-driving-cars-disengagement-rate/?guccounter=1

[EEVblog]

It's not the twisty crowded roads, it's the rain, hail and snow. The lidar is famously blinded by heavy precipitation.

I remember driving home one night from the city. It was raining hard and Sydney's streets were more than their usual nightmare with nighttime road works, huge flood lighting blinding everything, and contraflows galore that even I had a hard time figuring out. And I remember thinking that an autonomous car wasn't going to be able to handle this within my lifetime.
Maybe if it tailgated the car in front that would be a sensible AI tactic, but on it's own no freaking way.

[radar_macgyver]

If I were to guess, once a significant fraction of vehicles on the road are autonomous, there will be a push for integrating features into roads to help them navigate in a robust manner. Retro-reflectors for lidar or radar, inductive loop sensors or similar. This would make them safer to use during poor visibility conditions than unaided human drivers.

[Mr. Scram]

I remember driving home one night from the city. It was raining hard and Sydney's streets were more than their usual nightmare with nighttime road works, huge flood lighting blinding everything, and contraflows galore that even I had a hard time figuring out. And I remember thinking that an autonomous car wasn't going to be able to handle this within my lifetime.
Maybe if it tailgated the car in front that would be a sensible AI tactic, but on it's own no freaking way.

I have a hard time believing the human sensor array can't be topped by a carefully designed technological one for this kind of data acquisition.

[schmitt trigger]

I would like to see a self driving car in the streets of Rome. With kamikaze scooters cutting routinely in front of you in twisting roads.

For a Roman driver, he would simply honk the horn and hurl a pair of delicious Italian expletives, and he would continue driving.
For AI, it would trigger a scram reaction on the brakes.

[EEVblog]

I remember driving home one night from the city. It was raining hard and Sydney's streets were more than their usual nightmare with nighttime road works, huge flood lighting blinding everything, and contraflows galore that even I had a hard time figuring out. And I remember thinking that an autonomous car wasn't going to be able to handle this within my lifetime.
Maybe if it tailgated the car in front that would be a sensible AI tactic, but on it's own no freaking way.

I have a hard time believing the human sensor array can't be topped by a carefully designed technological one for this kind of data acquisition.

You can train it to do anything, but that's not the point.

[nctnico]

It's not the twisty crowded roads, it's the rain, hail and snow. The lidar is famously blinded by heavy precipitation.

I remember driving home one night from the city. It was raining hard and Sydney's streets were more than their usual nightmare with nighttime road works, huge flood lighting blinding everything, and contraflows galore that even I had a hard time figuring out. And I remember thinking that an autonomous car wasn't going to be able to handle this within my lifetime.
Maybe if it tailgated the car in front that would be a sensible AI tactic, but on it's own no freaking way.

IMHO you are falling in the same trap as many others: a self driving car can use different sensors so it can get more information than just the visual spectrum humans can detect. The whole discussion reminds me of an example of a machine which packs eggs in boxes. It doesn't use robotic fingers but vacuum and a suction cup. As a consequence it can pack eggs faster than any human can do.

Look at this example image from Flir which shows how a situation you describe looks using a thermal imaging camera:

[magic]

My guess is it will work in some rich American / West European areas where roads are reasonably maintained, poor and dumb people and their junk cars are kept out by economics, everybody is obsessed with obeying the law (which is not the case in most of the world) and there will be enough technology fanatics to pressure regulators into making accommodations for their toys if something doesn't work out.
There are such people, there are such politicians, there are such places. I can imagine SV being one - the terrain is flat, weather is mostly good, bio-drivers are rather timid, and the place is full of rich nerds. But now, my home city in Poland with constant roadworks and some bizarrely creative solution, or Warsaw with its notoriously asshole drivers - different story.

[Mr. Scram]

You can train it to do anything, but that's not the point.

I'm not taking about training, I'm talking about sensors. Humans have a fixed and limited array of them and cars can be fitted with anything we come up with. If that doesn't already provide an advantage, it most definitely will in the future. Human eyeballs and ears aren't that remarkable as means to drive a vehicle.

[apis]

All we ever see is these autonomous cars doing (relatively) basic stuff, as advanced and impressive as that is of course, again it's not anything close to what a human is capable of.

Well, the proof of the pudding is in the eating. We'll see.

Can't say I've seen one get out of the way of an ambulance. Pull over when the cops point their finger. Notice that someone is walking toward their car and is about to hop in and so wait for a minute.

I've heard them mention all of those scenarios saying they can handle that. The Waymo cars have microphones that actively listen for sirens, they can recognise emergency vehicles and they will take appropriate actions to let them pass. They use neural networks to predict if someone is about to get into (or out of) a car, and I've seen a video of the cars following the commands from a traffic police using hand gestures in an intersection.

Car being directed by traffic cop at an intersection with broken lights:


Car predicting intention of bicyclists, slowing down to let them pass:

While the cars don't rely on neural networks to determine if there is an obstacle on the road in front of the car (that is what they use the lidar for)

The Tesla doesn't have a LIDAR, and Musk recently said any company who uses one is doomed.

But the Tesla isn't a self driving car, it has some fancy cruise control (classified as level 2) and they kill people regularly. Tesla doesn't use lidar because it would be too expensive to put into all their cars. Everyone else does though, since they are not making cars but drivers (full level 4 autonomy). Waymo and GM develop their own in-house lidar systems.

It's not the twisty crowded roads, it's the rain, hail and snow. The lidar is famously blinded by heavy precipitation.

I remember driving home one night from the city. It was raining hard and Sydney's streets were more than their usual nightmare with nighttime road works, huge flood lighting blinding everything, and contraflows galore that even I had a hard time figuring out. And I remember thinking that an autonomous car wasn't going to be able to handle this within my lifetime.
Maybe if it tailgated the car in front that would be a sensible AI tactic, but on it's own no freaking way.

They can handle light rain, but in heavy rain they pull over and stop. That is clearly not a long time solution and it will probably be a while before they solve that problem. Since the lidar is blinded in that situation (and I don't see how they could fix that) they would have to rely only on cameras and radar, and as we have seen from Tesla, they can't drive safely with only cameras and radar. A lot of progress has been made on computer vision in the last couple of years though and vehicle radar is getting better as well. They will probably get there eventually, but it's hard to predict when, it could take 10 years or 30.

[EEVblog]

But the Tesla isn't a self driving car

Perhaps you haven't been reading the news. Telsa say they will have 1 million fully autonomous cars on the road next year.
That's the whole point of this thread.

[EEVblog]

I've heard them mention all of those scenarios saying they can handle that. The Waymo cars have microphones that actively listen for sirens, they can recognise emergency vehicles and they will take appropriate actions to let them pass. They use neural networks to predict if someone is about to get into (or out of) a car, and I've seen a video of the cars following the commands from a traffic police using hand gestures in an intersection.

Cute, but the world is one of infinite variability.
Last holidays I approached an intersection that had a guy off his face on drugs playing a raged traffic cop in the middle of an 8 lane highway. Laying down, banging on cars, chasing them, head butting the traffic lights, then the next minute directing traffic. Good luck AI.