Robo-Cab Predictions

[apis]

Yep, so you don't need perfect low latency connection in most cases, just tell the car that the traffic lights are broken and it will treat it as an intersection without traffic lights. What's nice is that all the other cars can also be given the same information, so when they reach the same broken traffic lights they already know what to do.

[nctnico]

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.

This is not a very strong argument. Who claimed self driving cars will wait forever in such a situation? One of the first things Waymo found out is that adhering strictly to the speed limit is unsafe so passing a red light in certain conditions is probably also part of the algorithm already.

And to paint a little bit of the future: With vehicle to vehicle communication chances are the self driving car is already aware of the broken traffic light and may even negotiate with other self driving cars on how to pass the crossing. It will be the human drivers causing the congestion.

[EEVblog]

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.

Won't happen, for obvious technical reasons, not to mention liability reasons.

[Mr. Scram]

Won't happen, for obvious technical reasons, not to mention liability reasons.

What obvious technical or liability reasons are there? There's no need for direct remote control and liability doesn't seem exceptionally complicated.

[EEVblog]

Won't happen, for obvious technical reasons, not to mention liability reasons.

What obvious technical or liability reasons are there? There's no need for direct remote control and liability doesn't seem exceptionally complicated.

I'm responding to the assertion that remote control by a human operator would be practical. It won't be, and if you can't think of reasons why that could be the case then you need to think about the problem a bit longer. This isn't a done flying over Afghanistan.

Liability is obvious. The first deadly accident caused by remote human operator will bring the whole idea crashing down and you won't be able to get liability cover for it.

[apis]

Remember, it won't be like remote control of a toy car. It will be telling the car they can safely ignore the malfunctioning traffic light, or telling the car it needs to make an u-turn and take another route. The car will still be driving autonomously, but it will get help with deciding how to deal with those rare unknowns. Latency or even loss of connection won't be much of an issue as long as it receive the new instructions.

[Mr. Scram]

I'm responding to the assertion that remote control by a human operator would be practical. It won't be, and if you can't think of reasons why that could be the case then you need to think about the problem a bit longer. This isn't a done flying over Afghanistan.

Liability is obvious. The first deadly accident caused by remote human operator will bring the whole idea crashing down and you won't be able to get liability cover for it.

We and others did think about it. That's where the Mars rover story comes in. You don't need to drive vehicles like a remote controlled car. You instruct them to take a certain action and the hardware figures out locally how to do that safely. An example would be to ignore the traffic lights so the intersection is treated as one without lights. Multiple automated vehicles could even negotiate who goes first amongst themselves. The idea is to leave the bulk of the work to automation and have humans take a look if and when the rare exception occurs. Liability doesn't seem complicated. All sorts of life and death decisions are made remotely. Bridges are raised remotely and fatal accidents do occasionally happen. People tend to prefer humans at the proverbal wheel, even if they're not physically present. Automated processes get a lot less leeway.

[nctnico]

Won't happen, for obvious technical reasons, not to mention liability reasons.

What obvious technical or liability reasons are there? There's no need for direct remote control and liability doesn't seem exceptionally complicated.

I'm responding to the assertion that remote control by a human operator would be practical. It won't be, and if you can't think of reasons why that could be the case then you need to think about the problem a bit longer. This isn't a done flying over Afghanistan.

Liability is obvious. The first deadly accident caused by remote human operator will bring the whole idea crashing down and you won't be able to get liability cover for it.

Why not? Usually car insurance is tied to the owner of the car and not the operator. It is not like deadly accidents aren't happening with people behind the wheel. Legally there is nothing new when it comes to self driving cars. It is up to the insurance companies to do the math on what premium they have to charge.

[maginnovision]

I'm responding to the assertion that remote control by a human operator would be practical. It won't be, and if you can't think of reasons why that could be the case then you need to think about the problem a bit longer. This isn't a done flying over Afghanistan.

Liability is obvious. The first deadly accident caused by remote human operator will bring the whole idea crashing down and you won't be able to get liability cover for it.

We and others did think about it. That's where the Mars rover story comes in. You don't need to drive vehicles like a remote controlled car. You instruct them to take a certain action and the hardware figures out locally how to do that safely. An example would be to ignore the traffic lights so the intersection is treated as one without lights. Multiple automated vehicles could even negotiate who goes first amongst themselves. The idea is to leave the bulk of the work to automation and have humans take a look if and when the rare exception occurs. Liability doesn't seem complicated. All sorts of life and death decisions are made remotely. Bridges are raised remotely and fatal accidents do occasionally happen. People tend to prefer humans at the proverbal wheel, even if they're not physically present. Automated processes get a lot less leeway.

The Mars Rover is an entirely different thing. there are no rules, and it's not terribly smart resulting in manual control in many situations. From what the Rover driver told me it's not even close to what is required here. It also takes a long time to make a move, find out it barely moved at all and then try again. For remote command of an earth car it actually needs to be aware of every safety rule and be able to manually disable any safety related check. Now let's consider the number of issues an hour and think about how that's handled. We've established these are cases the car failed and NEEDS manual control so how many external operators do we have? Is it really more efficient than people driving themselves? We could require more training for a license for drastically less than all the additional infrastructure for "self driving" cars. Another issue would be that now you have the ability to disable safety requirements from being met who is going to hack it first?

[apis]

Why not? Usually car insurance is tied to the owner of the car and not the operator. It is not like deadly accidents aren't happening with people behind the wheel. Legally there is nothing new when it comes to self driving cars. It is up to the insurance companies to do the math on what premium they have to charge.

Would Waymo even need insurance? They can count on Alphabet to cover any unexpected expenses for them.

[Mr. Scram]

The Mars Rover is an entirely different thing. there are no rules, and it's not terribly smart resulting in manual control in many situations. From what the Rover driver told me it's not even close to what is required here. It also takes a long time to make a move, find out it barely moved at all and then try again. For remote command of an earth car it actually needs to be aware of every safety rule and be able to manually disable any safety related check. Now let's consider the number of issues an hour and think about how that's handled. We've established these are cases the car failed and NEEDS manual control so how many external operators do we have? Is it really more efficient than people driving themselves? We could require more training for a license for drastically less than all the additional infrastructure for "self driving" cars. Another issue would be that now you have the ability to disable safety requirements from being met who is going to hack it first?

I'm not sure you're getting the point. Nobody is claiming the Mars rover could participate autonomously in Earth traffic. The point was how it deals with decision making, not having a real time operator and unusual situations. You don't need an Earth bound car to be able to deal with all exceptions. Hybrid solutions are viable and much more realistic than expecting a fully autonomous car right away. That's pretty much how much of our automation works. You wouldn't need that many human operators. Apparently the Waymo cars are currently at one intervention every 18000 kilometres.

[maginnovision]

The Mars Rover is an entirely different thing. there are no rules, and it's not terribly smart resulting in manual control in many situations. From what the Rover driver told me it's not even close to what is required here. It also takes a long time to make a move, find out it barely moved at all and then try again. For remote command of an earth car it actually needs to be aware of every safety rule and be able to manually disable any safety related check. Now let's consider the number of issues an hour and think about how that's handled. We've established these are cases the car failed and NEEDS manual control so how many external operators do we have? Is it really more efficient than people driving themselves? We could require more training for a license for drastically less than all the additional infrastructure for "self driving" cars. Another issue would be that now you have the ability to disable safety requirements from being met who is going to hack it first?

I'm not sure you're getting the point. Nobody is claiming the Mars rover could participate autonomously in Earth traffic. The point was how it deals with decision making, not having a real time operator and unusual situations. You don't need an Earth bound car to be able to deal with all exceptions. Hybrid solutions are viable and much more realistic than expecting a fully autonomous car right away. That's pretty much how much of our automation works. You wouldn't need that many human operators. Apparently the Waymo cars are currently at one intervention every 18000 kilometres.

I think you're underestimating how often manual operation is required for the rover. It's not an autonomous vehicle, it's movements are entirely dictated by humans. It doesn't make large journeys WITHOUT intervention because it's so simple and prone to getting off. Maybe the rover driver knows things you haven't read, I don't know. He could also be wrong, I don't know. How many vehicles do you expect to be on the road? Waymo is working in a small known area. Let's expand that to an unlimited area, and what... Millions of cars? Hundreds? Once the area is expanded to the whole planet you're going to have a lot more instances of "interventions" regardless of the number of vehicles. You also can't plan for average, especially when you start out, you plan for the worst case because otherwise there are going to be times where nobody goes anywhere because way too many require manual intervention than was planned. That happens a couple times and demand will pretty quickly go to 0.

[coppice]

Usually car insurance is tied to the owner of the car and not the operator.

That sounds bizarre. Insurance in most countries is tied to a combination of car and driver, and the rates vary massively between a teenager who just got their licence and a mature driver with a clean accident record. Its the driver who is in trouble if they operate a car they are not insured to drive.

[apis]

Waymo is working in a small known area. Let's expand that to an unlimited area, and what... Millions of cars? Hundreds? Once the area is expanded to the whole planet you're going to have a lot more instances of "interventions" regardless of the number of vehicles.

Actually this appears to be a common misunderstanding. That number of km/intervention is for their prototype testing in all of California. Waymo have been test driving all over USA (more than 16 million km / 10 million mi) it's not only in a small known area.

They have launched a pilot service in Phoenix that's open to the public, and that taxi service only operates in a limited area ("Waymo One is currently available in the East Valley of Phoenix Arizona, including Chandler, Gilbert, Mesa and Tempe."). The number of km/intervention in the Phoenix area is unknown but likely much better than their average in California. We only know that number for CA because everyone who's testing in CA is required by law to publish their disengagement rate.

[coppice]

Look at the talks by leading people working on autonomous cars on Lex Fridman's channel on YouTube. There are talk by people from most of the key players in autonomous cars there. None of them expect a true autonomous car to appear for a number of years. Their disagreements are more about whether its 10 years, 20 years or even further away. All they expect to see in the near future are cars operating in very constrained domains, in good weather, and with strict geo-fencing so they stay in well characterised places.

[coppice]

Waymo is working in a small known area. Let's expand that to an unlimited area, and what... Millions of cars? Hundreds? Once the area is expanded to the whole planet you're going to have a lot more instances of "interventions" regardless of the number of vehicles.

Actually this appears to be a common misunderstanding. That number of km/intervention is for their prototype testing in all of California. Waymo have been test driving all over USA, not only in a small known area.

That's not what Waymo's own people say in talks. They are using a lot more than one area for testing, but they are still very much geo-fenced. Even where they do operate, they stick to geo-fenced places they have characterised well, They can only function autonomously where in places where they have produced a high resolution map for lidar navigation, and they currently only have those for selected areas.

They have launched a pilot service in Phoenix that's open to the public, and that taxi service only operates in a limited area ("Waymo One is currently available in the East Valley of Phoenix Arizona, including Chandler, Gilbert, Mesa and Tempe.").

Its not open to the public. Its only available to a pool of registered people, to travel between selected points. Its being used for regular daily journeys, but its far from being a flexible public system right now

[apis]

Look at the talks by leading people working on autonomous cars on Lex Fridman's channel on YouTube. There are talk by people from most of the key players in autonomous cars there. None of them expect a true autonomous car to appear for a number of years. Their disagreements are more about whether its 10 years, 20 years or even further away. All they expect to see in the near future are cars operating in very constrained domains, in good weather, and with strict geo-fencing so they stay in well characterised places.

That mostly agrees with everything we've said. But did they really say "very constrained" though? Constrained for sure, but that's a big difference. And they don't need "good weather", they can't handle certain types of very bad weather at the moment, like heavy rain/snow.

My guess is that Waymo will launch their service in Phoenix (and possibly other cities) without safety drivers within the next five years. Then they will gradually expand the areas where they operate. It will probably take decades (the years quickly add up) before they can drive all over the US, not to mention the rest of the world. You have to crawl before you can walk, walk before you can run.

What they are referring to is full level 5 autonomy, which is indeed at least 10 years off, but you don't need level 5 autonomy to start a robo-cab service in some major cities.

[apis]

That's not what Waymo's own people say in talks. They are using a lot more than one area for testing, but they are still very much geo-fenced. Even where they do operate, they stick to geo-fenced places they have characterised well, They can only function autonomously where in places where they have produced a high resolution map for lidar navigation, and they currently only have those for selected areas.

"With Atlanta, Waymo has now officially tested in 25 cities across the U.S. That includes complex places such as the foggy hills of SF, the snowy streets of Michigan and the rainy roads of Kirkland, WA."

https://twitter.com/Waymo/status/955563835422687232

I've seen a much more detailed map, but can't find it now. They test outside those cities as well, all over the states they are allowed to test in (i.e. in the green areas). It's probably fair to say most of their efforts are concentrated to certain areas, but that's to be expected. If you give it some thought, it's obvious they would want to test in as many different situations as possible.

If by "small known area" you mean places where they have detailed maps, then sure, they use detailed maps as everyone knows, but it's not a "small area" by any definition.

They have launched a pilot service in Phoenix that's open to the public, and that taxi service only operates in a limited area ("Waymo One is currently available in the East Valley of Phoenix Arizona, including Chandler, Gilbert, Mesa and Tempe.").

Its not open to the public. Its only available to a pool of registered people, to travel between selected points. Its being used for regular daily journeys, but its far from being a flexible public system right now

Well, I don't live in Phoenix, but they write on their homepage:
"Waymo One is our commercial self-driving service that allows members of the public to travel from place to place in one of Waymo's self-driving cars.

Riders can use our app to call our self-driving vehicles 24 hours a day, 7 days a week. They can use it to get around several cities in the Metro Phoenix area. Whether it's for a fun night out or just to get a break from driving, our riders get the same clean vehicle every time and a driver with over 10 million miles of experience on public roads.

Over time, we plan to make Waymo One available to more members of the public as we add more vehicles and drive in more places."

They've recently partnered with Lyft as well:
"Six months ago, we launched Waymo One, our commercial self-driving ride-hailing service in the Metro Phoenix area. Over that time we’ve grown to serve over 1,000 riders who hail Waymo cars each day to commute to and from work, bring their kids to school, get to the grocery store, and even to avoid parking at trailheads before a big run.

Today, we’re pleased to share some details of our partnership with Lyft, which will help us welcome even more riders in Metro Phoenix to experience self-driving technology. As a first step, we’ll deploy 10 Waymo vehicles on Lyft over the next few months. Once Waymo vehicles are on the platform, Lyft users in the area will have the option to select a Waymo directly from the Lyft app for eligible rides."
https://medium.com/waymo/partnering-with-lyft-to-serve-more-riders-in-metro-phoenix-a9ce8709843e

[Mr. Scram]

I think you're underestimating how often manual operation is required for the rover. It's not an autonomous vehicle, it's movements are entirely dictated by humans. It doesn't make large journeys WITHOUT intervention because it's so simple and prone to getting off. Maybe the rover driver knows things you haven't read, I don't know. He could also be wrong, I don't know. How many vehicles do you expect to be on the road? Waymo is working in a small known area. Let's expand that to an unlimited area, and what... Millions of cars? Hundreds? Once the area is expanded to the whole planet you're going to have a lot more instances of "interventions" regardless of the number of vehicles. You also can't plan for average, especially when you start out, you plan for the worst case because otherwise there are going to be times where nobody goes anywhere because way too many require manual intervention than was planned. That happens a couple times and demand will pretty quickly go to 0.

We already established on two separate occasions that Waymo is reporting one intervention every 18000 kilometres. It'd also be obtuse to carbon copy the Mars rover setup to street cars rather than seeing it as the concept to base development on. At this point I'm starting to wonder whether people are wilfully ignoring or misinterpreting the conversation so far.

[apis]

We already established on two separate occasions that Waymo is reporting one intervention every 18000 kilometres. It'd also be obtuse to carbon copy the Mars rover setup to street cars rather than seeing it as the concept to base development on. At this point I'm starting to wonder whether people are wilfully ignoring or misinterpreting the conversation so far.

And to be clear, 18000 km per intervention is for their testing in the state of California. Everyone says it's so much easier to drive in Phoenix AZ (where it seldom rains), so one has to assume the number of km/intervention in the geo-fenced area there is significantly higher.

[maginnovision]

I think you're underestimating how often manual operation is required for the rover. It's not an autonomous vehicle, it's movements are entirely dictated by humans. It doesn't make large journeys WITHOUT intervention because it's so simple and prone to getting off. Maybe the rover driver knows things you haven't read, I don't know. He could also be wrong, I don't know. How many vehicles do you expect to be on the road? Waymo is working in a small known area. Let's expand that to an unlimited area, and what... Millions of cars? Hundreds? Once the area is expanded to the whole planet you're going to have a lot more instances of "interventions" regardless of the number of vehicles. You also can't plan for average, especially when you start out, you plan for the worst case because otherwise there are going to be times where nobody goes anywhere because way too many require manual intervention than was planned. That happens a couple times and demand will pretty quickly go to 0.

We already established on two separate occasions that Waymo is reporting one intervention every 18000 kilometres. It'd also be obtuse to carbon copy the Mars rover setup to street cars rather than seeing it as the concept to base development on. At this point I'm starting to wonder whether people are wilfully ignoring or misinterpreting the conversation so far.

I think you're missing my point. The rover isn't a metric that should be used in any way for self driving cars. That's the whole point of that. Don't even worry about it in the context of self driving cars. Waymo reports that for ONE area they drive. California isn't particularly known for extreme weather and hard to navigate roads, neither are those areas of arizona. Where is the kansas testing? New york city? Las vegas? We have poor comparisons for the rest of the US nevermind the world. That's my point, I'm not ignoring or misinterpreting anything. You also never answered how many cars you expect on the road and how many remote operators you see being necessary. It's a huge SCALE issue even if the rate of disengagements stays the same as the number of cars goes up. How do you expect they'd deal with hacking once you're required to add a way to disable safety features? What about bad actors as remote operators? Public perception is huge and there are many things that can outright end any chance of self driving vehicles gaining a foot hold in any market.

[Mr. Scram]

I think you're missing my point. The rover isn't a metric that should be used in any way for self driving cars. That's the whole point of that. Don't even worry about it in the context of self driving cars. Waymo reports that for ONE area they drive. California isn't particularly known for extreme weather and hard to navigate roads, neither are those areas of arizona. Where is the kansas testing? New york city? Las vegas? We have poor comparisons for the rest of the US nevermind the world. That's my point, I'm not ignoring or misinterpreting anything. You also never answered how many cars you expect on the road and how many remote operators you see being necessary. It's a huge SCALE issue even if the rate of disengagements stays the same as the number of cars goes up. How do you expect they'd deal with hacking once you're required to add a way to disable safety features? What about bad actors as remote operators? Public perception is huge and there are many things that can outright end any chance of self driving vehicles gaining a foot hold in any market.

The problem is that you seem so busy pointing out differences that you refuse to acknowledge the one part where to conceptual similarities line up. It's like refusing the acknowledge cows and geese can both feed on grass by pointing out one has wings and the other four feet and one is a mammal and the other isn't. Sure, but it's also missing the point.

The number of operators seems to be made into an issue where I'm not sure there actually is one. If you have one intervention every 16000 kilometres and the internet tells us cabs average 70000 miles or about 112000 kilometres a year that's 7 interventions per year per cab. That's supposing the technology doesn't advance. Let's be generous and say we assume one intervention takes an operator 30 minutes to resolve. Assuming about 250 working days in a year an operator can process over 4000 interventions in a year and therefore service close to 600 cabs. Note that there are no drivers in these cabs that need to be paid. Even if you pay operators double what taxi drivers earn and the numbers are off by a crazy margin you're still good. The scale doesn't appear to be an issue either as it's not as if hundreds of thousands of cabs appear on the roads overnight. We have various cars on the road already which are very vulnerable to hacking. The hacking ship has sailed.

https://www.wired.com/2015/07/hackers-remotely-kill-jeep-highway/

[David Hess]

PREDICTION #1:
Robocabs that can match a human driver in practicality are *many* years away. Perhaps even a decade or more.
Anyone talking next year is kidding themselves.
And they could still suck in many situations.

But they will become viable over limited, dense, and controlled urban routes much sooner.

PREDICTION #2:
People taking Robo-cabs won't like or trust traditional looking cars with an empty front seat and steering wheel whizzing around.
Expect something different to win out.

Johnny Cab!

PREDICTION #3:
Johnny Cab won't happen. People will detest humanoid robots trying to interact with them.

Put HAL on the dashboard.  What could possibly go wrong?

What about package delivery though where there are no passengers?

PREDICTION #4:
The "one bad experience" problem may seriously hamper initial Robo-Cab adoption.
And with social media how it is, all you'll hear about are the bad experiences.
The media & your Facebook friends will lap it up.
Although, Facebook could be dead by then...

I agree.  Image is everything.  Actual safety versus miles driven is irrelevant.

[vk6zgo]

Even though the continuing "accident toll" is an ongoing concern, the numbers of cars driven by humans are vast, as are the km/miles travelled, over many years, & in all sorts of road conditions.

The vast majority of the trips taken are safely completed.

On the other hand, there are a handful of self driving vehicles, some of which have already had accidents leading to death.

The argument is "We have operated this vehicle for X number of miles, which is more than the average driver does in Y time, hence, it is safer, statistically."

This reminds me of when germanium transistors first started to be widely used.

We were told "Accelerated life tests & statistics prove that these devices will have lifetimes far in excess of vacuum tubes----- of the order of twenty years!"

When they started to die after a few months to a year, we had to re-evaluate that information!

[apis]

On the other hand, there are a handful of self driving vehicles, some of which have already had accidents leading to death.

There has only been one fatality as far as I know, the Uber accident. The others are people not paying attention when using the Tesla autopilot which isn't a fully self driving car. You have to look at each self driving car project separately though, one can be perfectly safe and another can be a worse than a drunk driver. Only the safe ones should be allowed to drive on public roads.