Volvo bucks the industry, will sell LIDAR-equipped self-driving cars to customers by 2022

[Bombastinator]

2 hours ago, atxcyclist said:

I've seen computers crash and I've seen people crash, my issue lies with how a self-driving vehicle will prioritize safety/damage during a crash. Is it occupants first and screw everyone else? If there's a motorcyclist/bicyclist stopped in front of it at a stoplight, and someone hits the self-driver while slowing down for the light, dose the car cut the wheels hard and plow over the curb into whatever light poles and trashcans are up there, even if it would be more dangerous for the occupants of the car? Or is it just final destination for the rider? With a person driving, investigators (police) can assess fault and the expectation is that in the prior scenario, if there is time to react the driver behind would cut the wheel and make a line for a light pole to save the rider; Logically a driver would normally sacrifice their vehicle in order to not run someone over.

 

As someone that rides a motorcycle every day (I mean, not right now because of the human malware), I'd rather be looked at as a fragile object to a driver with empathy (probably), rather than a calculated variable in a computer program. It is 'artificial intelligence' after-all, no soul,  no actual freedom of choice, and no empathy.

I don’t see why human malware would prevent motorcycle riding.  It’s sort of by definition solitary.

[Sauron]

3 hours ago, atxcyclist said:

my issue lies with how a self-driving vehicle will prioritize safety/damage during a crash. Is it occupants first and screw everyone else? If there's a motorcyclist/bicyclist stopped in front of it at a stoplight, and someone hits the self-driver while slowing down for the light, dose the car cut the wheels hard and plow over the curb into whatever light poles and trashcans are up there, even if it would be more dangerous for the occupants of the car? Or is it just final destination for the rider? With a person driving, investigators (police) can assess fault and the expectation is that in the prior scenario, if there is time to react the driver behind would cut the wheel and make a line for a light pole to save the rider; Logically a driver would normally sacrifice their vehicle in order to not run someone over.

There are two issues with this:

1) the car can't recognize if the object it's about to hit is a person

2) the car can't predict exactly how much damage it will cause to itself and its surroundings by swerving off the road.

 

The most basic way to handle this is to avoid the situation altogether. Ideally the car should perceive the obstacle from far enough away that it's able to stop safely before hitting it; this is something a computer with an advanced sensor array should be able to do better than a human. At that point you've already reduced the dilemma to extremely rare edge cases where a human would be unlikely to react in time anyway.

 

At that point you can either accept the fatality and have the car strictly follow the law (which usually means hitting the anomalous object), knowing that you couldn't possibly account for all situations, or you can try to predict as many edge cases as possible and try to find a morally appropriate action for each one while also providing the car with means to detect and interpret these scenarios. The latter is, in many cases, simply beyond our current technology.

[kokakolia]

On 5/7/2020 at 5:29 AM, Bombastinator said:

So you’re saying that there are multiple methodologies for 3d mapping and lidar is only one, which has specific advantages and disadvantages.  I’m curious what the specific ones for lidar are over other different systems if you could possibly elaborate.

Oh boy! That could easily be a 2 hour seminar. But I'll try my best. 

 

LIDAR:

+ Quick

+ Can measure the ground through the vegetation. Great for topography maps

+ Easy setup

- Doesn't like water, reflections etc...

- Expensive AF

- Stores a lot of data, takes time to process

- Bulky

- Saps a lot of battery power

- Not the most precise. Expect 5cm precision at best. A lot of artifacts can happen

 

Photogrammetry:

(i.e. taking millions of photographs and stitching them together)

+ The quickest. You can shoot a photo every couple of milliseconds. 

+ Also used for topography maps. But vegetation is a problem.

+ Easiest to setup

- Stores a boatload of data (think terabytes), takes the longest to process. 

- Saps a lot of battery

- The least precise 

 

3D reflectorless: 

(i.e. manually using a total station to measure precise points)

+ The most precise with mm precision

+ Small data (just a few megabytes for a large job)

+ Easy to process since you have the minimum qty of points. All you have to do is connect the dots. And that is done automatically through a code library and CAD software.

- Line of sight required between control points and what you're measuring 

- Setting this up is tedious. You basically have to do an old-school traverse to setup precise control points with known coordinates. Well, that's true for any high-precision mapping. 

- S L O W. You have to aim and shoot every point, all the while entering codes and attributes. Also keep an eye on the instrument to make sure it's level. You can waste a day of work if the instrument isn't level. 

 

3D scan:

(i.e. telling the total station to fire millions of points in all directions to create a 3D scan)

+ Relatively quick. Just setup the total station, hit a button and wait for ~20 minutes

+ Still precise with cm precision

+ The bundled software will let you draw vectors out of sharp features. So you can end up with an output similar to 3D reflectorless. 

- Line of sight required between control points and what you're measuring

- Large data

- Consumes power (just carry more batteries in your vehicle, it's hardly a problem)

- You still have to setup the total station with control points everywhere. That's like the most critical and time-consuming part of the job. Screw this up and everything you measure is garbage. 

- Kinda slow because of the setup process

- Processing is still a lot more work than with 3D reflectorless (even with the software) because you have millions of points (VS hundreds). And you can't input codes and attributes on the points. 

 

GPS in the field:

+ cheapest solution

+ easiest to setup

+ relatively quick, you can measure points manually with attributes and codes 

+ Continuous topo measures points automatically for a specified interval of time. So you can stick your GPS antenna on a truck or quad, and measure the topography as you're driving around in a grid. I did a lot of that. 

+ Can enter codes and attributes, so the data output is small and easy to process like 3d reflectorless. 

+ lightweight. 

+ No line of sight required between control points and/or total station

- Not the most precise with ~3cm precision. 

- Not suitable for measuring building structures due to precision and the wide GPS antenna on top of the pole makes it so that you can't measure points on the foot of walls or ledges. You can always do offsets. But at the point you may as well use a total station with 3D reflectorless for buildings. 

- Works poorly or not at all close to walls or underneath trees. That's how you get weird measurement errors. Some points can be 5m out. The precision delta on the screen frequently lies. So you'll always have a few points which are way off. 

 

 

[Bombastinator]

13 minutes ago, kokakolia said:

Oh boy! That could easily be a 2 hour seminar. But I'll try my best. 

 

LIDAR:

+ Quick

+ Can measure the ground through the vegetation. Great for topography maps

+ Easy setup

- Doesn't like water, reflections etc...

- Expensive AF

- Stores a lot of data, takes time to process

- Bulky

- Saps a lot of battery power

- Not the most precise. Expect 5cm precision at best. A lot of artifacts can happen

 

Photogrammetry:

(i.e. taking millions of photographs and stitching them together)

+ The quickest. You can shoot a photo every couple of milliseconds. 

+ Also used for topography maps. But vegetation is a problem.

+ Easiest to setup

- Stores a boatload of data (think terabytes), takes the longest to process. 

- Saps a lot of battery

- The least precise 

 

3D reflectorless: 

(i.e. manually using a total station to measure precise points)

+ The most precise with mm precision

+ Small data (just a few megabytes for a large job)

+ Easy to process since you have the minimum qty of points. All you have to do is connect the dots. And that is done automatically through a code library and CAD software.

- Line of sight required between control points and what you're measuring 

- Setting this up is tedious. You basically have to do an old-school traverse to setup precise control points with known coordinates. Well, that's true for any high-precision mapping. 

- S L O W. You have to aim and shoot every point, all the while entering codes and attributes. Also keep an eye on the instrument to make sure it's level. You can waste a day of work if the instrument isn't level. 

 

3D scan:

(i.e. telling the total station to fire millions of points in all directions to create a 3D scan)

+ Relatively quick. Just setup the total station, hit a button and wait for ~20 minutes

+ Still precise with cm precision

+ The bundled software will let you draw vectors out of sharp features. So you can end up with an output similar to 3D reflectorless. 

- Line of sight required between control points and what you're measuring

- Large data

- Consumes power (just carry more batteries in your vehicle, it's hardly a problem)

- You still have to setup the total station with control points everywhere. That's like the most critical and time-consuming part of the job. Screw this up and everything you measure is garbage. 

- Kinda slow because of the setup process

- Processing is still a lot more work than with 3D reflectorless (even with the software) because you have millions of points (VS hundreds). And you can't input codes and attributes on the points. 

 

GPS in the field:

+ cheapest solution

+ easiest to setup

+ relatively quick, you can measure points manually with attributes and codes 

+ Continuous topo measures points automatically for a specified interval of time. So you can stick your GPS antenna on a truck or quad, and measure the topography as you're driving around in a grid. I did a lot of that. 

+ Can enter codes and attributes, so the data output is small and easy to process like 3d reflectorless. 

+ lightweight. 

+ No line of sight required between control points and/or total station

- Not the most precise with ~3cm precision. 

- Not suitable for measuring building structures due to precision and the wide GPS antenna on top of the pole makes it so that you can't measure points on the foot of walls or ledges. You can always do offsets. But at the point you may as well use a total station with 3D reflectorless for buildings. 

- Works poorly or not at all close to walls or underneath trees. That's how you get weird measurement errors. Some points can be 5m out. The precision delta on the screen frequently lies. So you'll always have a few points which are way off. 

 

 

Cool! Thanks.  I appreciate that.

 

from the way it sounds there is lidar, “total” which can be used two ways, photos, and gps. For cars “total” is out, and possibly photos as well.  It would be at least highly difficult.  lidar is a lot more accurate than gps which seems to have occasional deep flaws, but it’s got its own issues.  I can see why lidar is looked at with hope for getting past level 3.  Not sure it would be enough though.

[pzspah]

Volvo are China owned so no thanks regardless of cool tech. 

[kokakolia]

6 hours ago, Bombastinator said:

Cool! Thanks.  I appreciate that.

 

from the way it sounds there is lidar, “total” which can be used two ways, photos, and gps. For cars “total” is out, and possibly photos as well.  It would be at least highly difficult.  lidar is a lot more accurate than gps which seems to have occasional deep flaws, but it’s got its own issues.  I can see why lidar is looked at with hope for getting past level 3.  Not sure it would be enough though.

Perhaps the solution for level 3 is very dystopian, in a sense where you remove all of the randomness caused by living beings. You can achieve that by only allowing self-driving cars on the road. And then you barricade the roads from people and animals.

 

I just literally described an automated underground train . Yep it already exists and it's awesome. 

 

I'm just not thrilled with getting past level 3. For me it's just a means for Uber to increase profits by cutting the drivers. I'm all for public transportation. 

[atxcyclist]

9 hours ago, Bombastinator said:

I don’t see why human malware would prevent motorcycle riding.  It’s sort of by definition solitary.

Up until very recently here, you were risking a $1000 fine or jail time if your couldn't justify the reason for your journey. To and from work if you were essential or going to the store or bank was fine, but just riding around wasn't covered.

[Bombastinator]

34 minutes ago, atxcyclist said:

Up until very recently here, you were risking a $1000 fine or jail time if your couldn't justify the reason for your journey. To and from work if you were essential or going to the store or bank was fine, but just riding around wasn't covered.

Ah.  State dependent.  Wouldn’t have happened where I am.

[BuckGup]

On 5/6/2020 at 9:57 AM, kokakolia said:

George Hotz figured out self-driving cars many years ago. He can turn a car driverless for a thousand bucks with off the shelf tech and his own software. It's a shocker that not all new cars are driver-less at this point. Here's a documentary explaining it all:

He didn't really figure it out. He had a computer watch him drive for many hours and then kept training the model to learn objects. His system does not work at all in snowy conditions, dark roads, rain, and fog. It's a cool concept but there is a reason you need to use sonar, lidar, and ultrasonic senors. You can't solely rely on vision. Even as humans we are fooled a lot by our vision 

[dalekphalm]

On 5/6/2020 at 11:19 AM, Sauron said:

The price in the article is a little misleading - it's true that the highest end LIDARs cost tens of thousands but you probably don't need anything nearly as precise for a car. For comparison, I've worked with a Leica LIDAR that made point clouds of tens of millions of points with incredible accuracy at a range of over 50 metres and that "only" cost $30k. A car doesn't need more than a few thousand points and you can get away with covering a fraction of the y axis that the Leica sensor would. More affordable industrial sensors used to pilot, for example, a robotic forklift would only cost the manufacturer a few thousand and are probably plenty for a car.

 

Also, as far as I know, all "self driving" cars have a Lidar on board. It's that spinning thing you can see on the top here:

not all cars have it on display like this but it's necessary for close range object detection. A Radar isn't precise enough to tell you the shape of an object.

Just an addendum to this, most self driving cars use Lidar.

 

Tesla doesn't, though. They rely purely on cameras and a few other onboard sensors like the front facing Radar. The Tesla method focuses on making the AI smart enough that it can look at the camera feeds exactly the same way a human could - but better (since it has pretty much 360 degree camera coverage, plus the benefit of reflexes no human could ever achieve).

 

Which method is better? Only history will tell us. Both have pros and cons. Lidar is more expensive, and relies on moving parts, which means significantly higher chance of failure. But Lidar can see things that a regular camera cannot - it can also more accurately map 3D environments, etc.

 

The camera method is cheaper, and can still 3D map the environment by using multiple cameras at different angles. And depending on the camera used, you can still achieve "better than human" vision, that can see through certain obstacles (disclaimer, I don't think Tesla is using non-standard camera tech - but fundamentally, they could if they wanted).

 

Ultimately I think both methods can achieve full self driving in time.

[YellowJersey]

I just hope that the types of lidar that fry camera sensors don't become widespread

[Bombastinator]

25 minutes ago, YellowJersey said:

I just hope that the types of lidar that fry camera sensors don't become widespread

..or start appearing on cats..

https://www.youtube.com/watch?v=e5fiBFhf9OQ

[Sauron]

9 hours ago, dalekphalm said:

Tesla doesn't, though. They rely purely on cameras and a few other onboard sensors like the front facing Radar. The Tesla method focuses on making the AI smart enough that it can look at the camera feeds exactly the same way a human could - but better (since it has pretty much 360 degree camera coverage, plus the benefit of reflexes no human could ever achieve).

Interesting, that might be why they're falling behind WayMo and why their cars thought a truck was the sky

 

Anyway as Lidar technology becomes cheaper I don't see why they wouldn't integrate it - it gives you so much more information than any camera could. Also I would argue safety should be the priority here over cost, after all these are luxury cars to begin with.

9 hours ago, dalekphalm said:

The camera method is cheaper, and can still 3D map the environment by using multiple cameras at different angles.

Welllll in theory you can do that but that entirely relies on the AI being able to tell exactly where each object begins or ends. Even humans don't always get it right. A Lidar gives you almost perfect depth perception with little room for error and used in combination with cameras it gets rid of a lot of ambiguity.

9 hours ago, dalekphalm said:

depending on the camera used, you can still achieve "better than human" vision, that can see through certain obstacles (disclaimer, I don't think Tesla is using non-standard camera tech - but fundamentally, they could if they wanted).

What you gain in overall perception is often lost in processing because, as much as Tesla doesn't want to admit it, AI technology is still way behind the human brain. It can do select things better but not many others. For this reason you need as much perception as possible to make up for that loss.

9 hours ago, YellowJersey said:

I just hope that the types of lidar that fry camera sensors don't become widespread

Most Lidars don't do that. At some point I guess we'll need some regulation to that effect though.

[kokakolia]

@SauronThinking about electric cars, LIDAR seems counterproductive by using a lot of energy. The range will no doubt suffer. 

 

[Bombastinator]

5 minutes ago, kokakolia said:

@SauronThinking about electric cars, LIDAR seems counterproductive by using a lot of energy. The range will no doubt suffer. 

 

“A lot” is still a vague term.  Cars use “a lot” of energy just to move.  Everything is relative.

There’s an old trick with gas engine cars where you can use the starter battery and the starter to run the car for a short distance without using the engine. 
 

The same might apply to using the some light or playing the radio.  The question is how much compared to what the capacity and range of the car is. 

[Sauron]

8 minutes ago, kokakolia said:

@SauronThinking about electric cars, LIDAR seems counterproductive by using a lot of energy. The range will no doubt suffer. 

That's a problem with electric cars, not with LIDARs. Battery technology needs to get a lot better before auto EVs can become the norm. I'll take a "normal" electric car over having unsafe autonomous driving because the manufacturer couldn't power the LIDAR and offer a reasonable range.

[Bombastinator]

3 minutes ago, Sauron said:

That's a problem with electric cars, not with LIDARs. Battery technology needs to get a lot better before auto EVs can become the norm. I'll take a "normal" electric car over having unsafe autonomous driving because the manufacturer couldn't power the LIDAR and offer a reasonable range.

That argument has been done and done for over a hundred years.  A 300 mile range is a 300 mile range.  The math got done on this one many many years ago.  It’s about energy density.  Which is energy vs weight.  Gasoline has an energy density too.  It’s higher than LiIon batteries, but battery powered cars don’t need to carry a 400lb engine around to be able to access it.  The result is you can make a gas car go farther than an electric car, but a car only has to go as far as a car needs to go.  There were functional (kind of) electric cars way back in the 1970’s that used lead acid batteries which had a much worse energy density.  Electric cars were made that used batteries with even worse energy densities.  Jay Leno has an electric car over a hundred years old that originally used glass bottles filled with alkali as batteries.   There’s a video on the thing somewhere.  He replaced them with a bank of lead acid car batteries and apparently the thing worked even better.

[Sauron]

22 minutes ago, Bombastinator said:

That argument has been done and done for over a hundred years.  A 300 mile range is a 300 mile range.  The math got done on this one many many years ago.  It’s about energy density.  Which is energy vs weight.  Gasoline has an energy density too.  It’s higher than LiIon batteries, but battery powered cars don’t need to carry a 400lb engine around to be able to access it.  The result is you can make a gas car go farther than an electric car, but a car only has to go as far as a car needs to go.  There were functional (kind of) electric cars way back in the 1970’s that used lead acid batteries which had a much worse energy density.  Electric cars were made that used batteries with even worse energy densities.  Jay Leno has an electric car over a hundred years old that originally used glass bottles filled with alkali as batteries.   There’s a video on the thing somewhere.  He replaced them with a bank of lead acid car batteries and apparently the thing worked even better.

Yeah, so... we need better batteries.

[Bombastinator]

Just now, Sauron said:

Yeah, so... we need better batteries.

Sure. We will always need better batteries.  Right now we’re on the edge of making batteries with high enough energy density to be used by commercial aircraft.   We were stuck at lead acid for many years, then they figured out how to deal with lithium ion dendrite buildup and boom.  Advancement.  Right now the stumbling block is solid metallic lithium batteries which promise better stability and energy density.  There are two popular avenues for attack on that one currently.  Might happen real real soon, might not.  Research is like that.

[ARikozuM]

Seeing that the week after quarantine we've already had about 7 accidents on our 2-lane bridge, I'm inclined to trust a computer more than people. 

[ARikozuM]

6 minutes ago, Sauron said:

Yeah, so... we need better batteries.

I'd take a battery over transport of gasoline and a paltry <30% efficiency. 

[Sauron]

11 minutes ago, ARikozuM said:

I'd take a battery over transport of gasoline and a paltry <30% efficiency. 

Yeah, but not if it dies halfway through a trip and requires me to wait hours for it to be full again.

[ARikozuM]

37 minutes ago, Sauron said:

Yeah, but not if it dies halfway through a trip and requires me to wait hours for it to be full again.

Depends on how your route is. Best choice would be free fast charge stations for while you eat or rest and paid replacements for grab-n-go. Investment needs to be taken away from gas stations and put into EV stations. Granted, here in Florida we have free charging stations and Tesla superchargers everywhere. It's not like most cars will do excessively more than 300-350 miles on a tank. I also wouldn't consider long commutes common unless you're a rural/farm worker or truck driver, in which case we have diesels for those. 

[Sauron]

1 minute ago, ARikozuM said:

I also wouldn't consider long commutes common unless you're a rural/farm worker or truck driver, in which case we have diesels for those. 

Daily, no - but I wouldn't want to buy a car and then have it let me down precisely when it would be significantly more convenient than public transport or cycling, meaning medium/long road trips to areas that aren't densely populated. This partially depends on where you live but here if you want to go hiking or climbing a car is pretty much your only option.

[dalekphalm]

6 hours ago, Sauron said:

Daily, no - but I wouldn't want to buy a car and then have it let me down precisely when it would be significantly more convenient than public transport or cycling, meaning medium/long road trips to areas that aren't densely populated. This partially depends on where you live but here if you want to go hiking or climbing a car is pretty much your only option.

How frequently does the average driver actually go beyond 300 miles though?

 

Most? Likely a few times a year. In those cases? Just rent a long range vehicle - the cost savings of charging an EV vs buying fuel will more than pay for the rental and the fuel required for the trip.