GM's Cruise teases it's 2019 autonomous car

cros13

Platform based on the Chevrolet Bolt.
A few thousand will be produced for large scale testing.
Some will see service as part of public trials in 2019.

https://www.youtube.com/watch?v=MvP82IsGqNc

ted1

Why not use rear facing seats with a table in the middle ?

Dont be at yourself

ted1 wrote: »

Why not use rear facing seats with a table in the middle ?

Indeed, a huge lack of imagination on GM's part.

unkel

Old thread, but might as well put in this update as it's pretty relevant:

GM (or more specifically their autonomous driving division Cruise) has lobbied NHTSA (the US National Highway Traffic Safety Administration) to get permission to take the steering wheel out of 2,500 fully self driving cars. It is expected they will get the go ahead some time this year.

I presume this means that we have finally made the step from testing autonomous driving level 4 cars (where there still is a human supervising and taking control when needed) to autonomous driving level 5 cars (humans are no longer required)

ELM327

I was just pondering this the other day.
This is where I think Tesla's model of cameras will struggle vs lidar.
The step from lvl4 to lvl5 autonomy.

I presume these GM cars are using Lidar with the rotating detectors and cost many many thousands?

unkel

GM are using lidar. Came across an article the other day that suggested lidar mass produced might only cost a 3 figure sum per car and not the many thousand it is costing today

Not sure about whether lidar is the way forward. I'm tending to side with Musk in that it is overkill and that cameras will be more than adequate. Time will tell.

ELM327

As a tesla fanboy/cultist/whatever else you're having I usually drink the kool aid but I think Lidar is needed here.
For level 4 cameras and minimal radar will be fine but I just don't think they will be able to get to level 5 based primarily on cameras. Even the raven loaner I had, the AP was so bad in any poorly lit conditions. If it had lidar or better tech it would have been easily able to cope.

I guess we'll have to wait and see.

KCross

unkel wrote: »

GM are using lidar. Came across an article the other day that suggested lidar mass produced might only cost a 3 figure sum per car and not the many thousand it is costing today

Not sure about whether lidar is the way forward. I'm tending to side with Musk in that it is overkill and that cameras will be more than adequate. Time will tell.

Its hard to tell alright how it will pan out but my thinking is that if Lidar was the end game it wouldnt still be geo-fenced like it is today. They've had years at perfecting it and its still only working in geo-fenced areas. I'm sure it will be super reliable in those areas but thats not the end game. It has to work everywhere... thats the end game.

If Telsa can crack the camera/radar angle and the elusive edge cases it should win out as it will work everywhere, but still too many ifs there to be sure.

Maybe there is room for both tech's?

unkel

ELM327 wrote: »

For level 4 cameras and minimal radar will be fine but I just don't think they will be able to get to level 5 based primarily on cameras.

Below is more philosophical than based on current state of tech, but:

Humans only use "cameras" and are reasonably safe drivers, even though their observation and attention is far from perfect and neither is their processing power or their database of situations (experience)

I can see autonomous cars with cameras doing significantly better than humans (Tesla AP already was a few years ago). Once that has been established with statistics, isn't that more than adequate safety for allowing level 5 autonomous driving?

ELM327

unkel wrote: »

Below is more philosophical than based on current state of tech, but:

Humans only use "cameras" and are reasonably safe drivers, even though their observation and attention is far from perfect and neither is their processing power or their database of situations (experience)

I can see autonomous cars with cameras doing significantly better than humans (Tesla AP already was a few years ago). Once that has been established with statistics, isn't that more than adequate safety for allowing level 5 autonomous driving?

Tesla AP is great but it's not there yet. Neither my AP1 car or the AP3/HW3 loaner I had were good for anything except driving on well marked lanes with no left turnoff lanes.


From what I can see they are still level 2 systems. Hands off assist systems. And again, I stress, I'm a fanboy, I'm not a short to bash Tesla, they are the best system out there now by a mile but it's nowhere near ready for level 4 let alone level 5.



There is a case to be made to designate specific roads as level 4 (eg motorways) but other than that it's just not ready.


Finally - humans with a 0.1 crash rate per 10k km is much better perceived than AP with 0.01 crash rate in the same distance.

KCross

unkel wrote: »

GM are using lidar. Came across an article the other day that suggested lidar mass produced might only cost a 3 figure sum per car and not the many thousand it is costing today

Another thought just struck me.... Tesla have said that Lidar is a fools errand because of the cost not because its **** tech. If the cost did actually come down to a 3 figure sum then that would be a good development and maybe they could merge the tech's and have cameras and lidar and radar. Now that would be cool.

Tesla would have the competitive advantage as cracking camera AI is the hard piece and I think Tesla are the only ones trying it at level 5?

unkel

ELM327 wrote: »

Tesla AP is great but it's not there yet. Neither my AP1 car or the AP3/HW3 loaner I had were good for anything except driving on well marked lanes with no left turnoff lanes.


From what I can see they are still level 2 systems.

Of course they are only level 2 still. But my point is that once level 4 cars can be shown (with accident statistics) to be significantly safer than humans (say a factor 3 times safer at least), isn't that justifications to legally allow them to be level 5 cars, as requested by Cruise?

ELM327

unkel wrote: »

Of course they are only level 2 still. But my point is that once level 4 cars can be shown (with accident statistics) to be significantly safer than humans (say a factor 3 times safer at least), isn't that justifications to legally allow them to be level 5 cars, as requested by Cruise?

I'm not the decision maker or stakeholder in that signoff but if I were I wouldnt be granting it unless it was significantly better than currently available Tesla AP3

unkel

ELM327 wrote: »

I'm not the decision maker or stakeholder in that signoff but if I were I wouldnt be granting it unless it was significantly better than currently available Tesla AP3

Of course. I was talking test vehicles that have been tested at autonomous driving level 4 for millions of miles already (with little to no human intervention), like the Cruise test cars.

Not autonomous driving level 2 production cars like the Tesla AP3

kirving

KCross wrote: »

Another thought just struck me.... Tesla have said that Lidar is a fools errand because of the cost not because its **** tech. If the cost did actually come down to a 3 figure sum then that would be a good development and maybe they could merge the tech's and have cameras and lidar and radar. Now that would be cool.

Tesla would have the competitive advantage as cracking camera AI is the hard piece and I think Tesla are the only ones trying it at level 5?

Lidar is extremely expensive for lost of reasons, mostly the optic as far as I know, but so are high resolution cameras. High three figures per car at minimum these days. Something is needed to fill the gap between Radar for long range and low res cameras / ultrasonics at close range.

Cameras don't work at night, and infrared cameras are an entirely different technology to vision cameras so for now are again extremely expensive.

As it stands, a number of companies are making great headway into solid state lidar rather than using spinning sensor to cover a wide field of view.

Tesla, right now, are good enough, in the sense that your average driver is good enough. Their sensor suite is far from the most advanced, but the question is, does it need to be to satisfy future regulation? Traditional OEMs have a lot of experience with regulation, long term reliability and safety, and they're taking each step very cautiously - but then again so did Nokia when Samsung knocked at the door.

If you look at it another way with respect to autonomous taxi or shared vehicle services, it's easy to justify a €20k sensor suite when you don't need to a pay a taxi driver.

Lumen

I don't think LIDAR vs cameras is important, it's all just data.

The real issue is whether a silicon+software neural net can be trained to drive a car at speed in a way that isn't terrifying and dangerous, by anticipating the actions of humans and correcting for their mistakes. That's the essence of safe driving in a mixed human-machine environment, it's a "theory of mind" problem.

ELM327

Lumen wrote: »

I don't think LIDAR vs cameras is important, it's all just data.

The real issue is whether a silicon+software neural net can be trained to drive a car at speed in a way that isn't terrifying and dangerous, by anticipating the actions of humans and correcting for their mistakes. That's the essence of safe driving in a mixed human-machine environment, it's a "theory of mind" problem.

I believe it can, I think the big jump (apart from level 0 to level 2) is level 4 to 5. Hands off is a long way away, but monitored 100% hands off is not too far away.

Lumen

ELM327 wrote: »

I believe it can, I think the big jump (apart from level 0 to level 2) is level 4 to 5. Hands off is a long way away, but monitored 100% hands off is not too far away.

But if I have to pay attention I may as well drive myself. It's a confusion of responsibilities and a recipe for disaster.

Automation is great when it's complementary, like ABS or cruise control.

Humans are terrible at maintaining attention without interaction.

ELM327

Lumen wrote: »

But if I have to pay attention I may as well drive myself. It's a confusion of responsibilities and a recipe for disaster.

Automation is great when it's complementary, like ABS or cruise control.

Humans are terrible at maintaining attention without interaction.

I disagree, even with L2 autonomy (tesla autopilot) 500-600+km in a day feels like a breeze whereas driving a car with cruise control only I'd be mentally tired.


Hands off (but still requiring a human behind a wheel, just in case) would be great.

Lumen

ELM327 wrote: »

I disagree, even with L2 autonomy (tesla autopilot) 500-600+km in a day feels like a breeze whereas driving a car with cruise control only I'd be mentally tired.

Hands off (but still requiring a human behind a wheel, just in case) would be great.

But you're an experienced driver. The next generation will be permanent noobs. Do you really want noobs taking control in the middle of a emergency situation?

unkel

The next generation won't have to worry about it as their cars will be level 5 autonomous

You have a point though, already obvious in aviation. A younger flight simulator trained pilot would not have been able to land that commercial airliner on the Hudson river with no loss of life and no serious injuries like the old skool pilot did.

kirving

Lumen wrote: »

I don't think LIDAR vs cameras is important, it's all just data.

It's very important. Each have their strengths and weaknesses, but there needs to be a level of practicality whereby the technology becomes affordable to the general public, yet capable enough to work in all scenarios.

Right now, there are a number of companies who have a €500k+ cars which drive you from A to B in perfect safety, but in only a limited number of places where excellent data already exists of the environment.

The neural network part of the equation brings efficiencies essential to the success of autonomous cars of, and means the car isn't just iteratively running "can I stop in the distance I can see to be clear, even if that pedestrian jumps off the motorway bridge" which would be ridiculously impractical.

As you said, the interaction with other road users is key, and so the predictive, machine learning elements of the software are key to the success, but that learning relies on exceptional quality data being available at all times.

Lumping it all together as just "data", belies the complexity required of each sensor to gather useful data come hail, rain or shine.

Low quality data lumped together to train neural network would be a like a person who's going slowly blind saying "sure I know the road".

Dont be at yourself

I recommend setting aside some time to watch Tesla's presentation to investors on their autonomy programme, which covers cameras vs LIDAR (as well their broader business model plan of robotaxis etc).

https://www.youtube.com/watch?v=nou8ZTsf4_c

The biggest question for any product in tech is "Can it scale?", and it seems to me that a camera + neural net system is much more scaleable than LIDAR; in terms of cost, deployment, ability to adapt to freak situations etc. Not sure if it'll be good enough, or over what timeframe (and I haven't ordered FSD), but it's the horse I'd back.

ted1

Are cameras better or worse than the human eye? With night vision and car head lights is a camera that doesn’t get distracted or tired any worse than a human eye.

Lidar may be better, but it may be overkill

KCross

Good technical comparison between Lidar and Camera/Radar/Ultrasonic shown below.

It also talks about the recent Uber crash where the pedestrian was killed eventhough they had Lidar and it had "seen" the pedestrian... software deficiency rather than a technological one!