Sensors (overflow from 2018.39 thread)

[rnortman]

Moderator note (bmah): The first six posts in this thread were moved from the Software Update 2018.39 thread, as they're discussions aspects of blind-spot monitoring and sensor technology that are not specific to any given software version.

I think it boils down to do you trust the car or not. I understand the skepticism. If you think the car's sensors and computer cannot be trusted as good as a person then you will probably always be skeptical of level 3 autonomy and you will want to do a "shoulder check" to make sure the car is doing the right thing. But if and when the car's sensors and computers are good enough to do things like blind spot checks, then you won't need to worry. At that point, you can let the car do auto lane changes on its own.

"If and when the car's sensors ... are good enough" is the problem here. Using radar for blind spot monitoring, including detection of vehicles fast-approaching from the rear, is a tried-and-true and very inexpensive off-the-shelf technology that Tesla completely eschewed in favor of unproven, undeveloped vision systems, optimistically assuming that eventually they'd figure it out. I'd trust it right now if it had radar and/or lidar, but without that I may never trust it; we'll see how well it does in rain, snow, with dirty lenses, and with fast-moving traffic. No matter how good it gets, it would be better with a radar as a redundant system, and we all would have had actual blind spot monitoring for the past 2 years without waiting for v9. Since you can buy a $30k car with rear corner radar for blind spot monitoring, why do our $90k cars not have it? Elon's Iron Man ego is the reason, and you can bet there have been avoidable accidents over the past two years as a result.

BTW I will always be skeptical of L3 autonomy, but I am not skeptical of L4 autonomy. L3 makes very little sense; by the time you can do L3 safely you can do L4. If you can't do L4, then your L3 system is a disaster waiting to happen... and probably not waiting for very long.

 

[ItsNotAboutTheMoney]

"If and when the car's sensors ... are good enough" is the problem here. Using radar for blind spot monitoring, including detection of vehicles fast-approaching from the rear, is a tried-and-true and very inexpensive off-the-shelf technology that Tesla completely eschewed in favor of unproven, undeveloped vision systems, optimistically assuming that eventually they'd figure it out. I'd trust it right now if it had radar and/or lidar, but without that I may never trust it; we'll see how well it does in rain, snow, with dirty lenses, and with fast-moving traffic. No matter how good it gets, it would be better with a radar as a redundant system, and we all would have had actual blind spot monitoring for the past 2 years without waiting for v9. Since you can buy a $30k car with rear corner radar for blind spot monitoring, why do our $90k cars not have it? Elon's Iron Man ego is the reason, and you can bet there have been avoidable accidents over the past two years as a result.

BTW I will always be skeptical of L3 autonomy, but I am not skeptical of L4 autonomy. L3 makes very little sense; by the time you can do L3 safely you can do L4. If you can't do L4, then your L3 system is a disaster waiting to happen... and probably not waiting for very long.

Tesla is all-in on producing a cheap autonomous system. It isn't going to spend money on additional, sensors that add to hardware and manufacturing costs of that autonomous system. It's the same reason why there isn't a dedicated sensor for windshield wipers.

 

[X Fan]

Tesla is all-in on producing a cheap autonomous system. It isn't going to spend money on additional, sensors that add to hardware and manufacturing costs of that autonomous system. It's the same reason why there isn't a dedicated sensor for windshield wipers.

“Cheap” or a system that is cost efficient but appropriate for the end user?

 

[ItsNotAboutTheMoney]

“Cheap” or a system that is cost efficient but appropriate for the end user?

Cheap. It's going long.

For an autonomous vehicle, every dependency needs redundancy. That multiplies costs and complexity.

 

[croman]

It's always great when a manufacturer sells a $100k car and cheaps out on $50 in parts that greatly increases safety. Really should be lauded. /s

 

[rnortman]

Cheap. It's going long.

For an autonomous vehicle, every dependency needs redundancy. That multiplies costs and complexity.

First of all, automotive radar is cheap, relative to the cost of these cars, and getting cheaper -- as long as we're "going long". They could always drop the radar for cost savings after the make it work with vision.

Re: redundancy -- if they have both radar and vision, and either gives them adequate blind spot monitoring, then they are redundant already without requiring additional redundancy. And when both work, you have sensor diversity which makes everything better. Do you really think they're going to turn 8 cameras into 16 to satisfy redundancy? Plus fully redundant GPU capacity to process all those cameras? Much cheaper and simpler to add radar for redundancy.

But as is apparent in the design of the current generation of Teslas, redundancy was only very minimally considered. These cars will never be certified for fully driverless operation.

 

[ItsNotAboutTheMoney]

First of all, automotive radar is cheap, relative to the cost of these cars, and getting cheaper -- as long as we're "going long". They could always drop the radar for cost savings after the make it work with vision.

And Tesla could have added added adaptive cruise control the same way every other company did it. It added it as part of Autopilot.

Re: redundancy -- if they have both radar and vision, and either gives them adequate blind spot monitoring, then they are redundant already without requiring additional redundancy. And when both work, you have sensor diversity which makes everything better. Do you really think they're going to turn 8 cameras into 16 to satisfy redundancy? Plus fully redundant GPU capacity to process all those cameras? Much cheaper and simpler to add radar for redundancy.

Yes, I think that Tesla's plan is to get Autopilot to a more advanced level before it will add the hardware redundancy necessary to go autonomous. There's no commercial value to adding hardware until it's needed.

Radar can't do everything that vision can do. So, it wouldn't provide redundancy.

But as is apparent in the design of the current generation of Teslas, redundancy was only very minimally considered. These cars will never be certified for fully driverless operation.

The current hardware certainly won't do L5. But that's irrelevant. Tesla isn't adding specific hardware because it expects Autopilot to become capable of providing that functionality.

 

[rnortman]

And Tesla could have added added adaptive cruise control the same way every other company did it. It added it as part of Autopilot.

But... you do realize that they did add adaptive cruise control the same way over other company did it? They use an off the shelf forward radar and they process it in more or less the same way everybody else does. In AP1 TACC was exactly just an off the shelf adaptive cruise control system. In AP2 they futzed with it a bit but it still operated on exactly the same principles. Very slowly over time they've been tweaking it and making it use vision more, but it's still primarily a radar-based ACC.

Yes, I think that Tesla's plan is to get Autopilot to a more advanced level before it will add the hardware redundancy necessary to go autonomous. There's no commercial value to adding hardware until it's needed.
[...]
The current hardware certainly won't do L5. But that's irrelevant. Tesla isn't adding specific hardware because it expects Autopilot to become capable of providing that functionality.

But... they specifically sold these cars as having all the hardware needed for fully autonomous operation... are you saying that was fraud? Will they give us all new cars once they've developed the technology?

Radar can't do everything that vision can do. So, it wouldn't provide redundancy.

Radar can very, very easily do blind spot monitoring. It is very good at that. It would provide redundancy for this critical feature. (Critical in an L3+ vehicle anyway... non-critical in L2.)

 

[d21mike]

Moderator note (bmah): The next 14 posts in this thread were moved from a slightly different discussion in the 2018.39 thread.

I would say that v9 takes us to about 90% of the on ramp to off ramp feature. With v9, we are really really close now.



I used to think that Tesla was pursuing two separate paths, one for EAP and one for FSD. But I have changed my mind on that. I now think that Tesla is working one single path towards FSD and EAP will simply gain new features and gain competency over time until it eventually becomes FSD. So in that sense, I think V9 definitely moves us one step closer to FSD.

I did not go back and read this thread about AUDI A8 L3. BUT, is the New AUDI A8 really L3 based on the definitions in THIS Thread? If it is L3 is it because it allows the driver to do other things until the driver is needed even though this only works (I believe) in heavy traffic going < 40mph? Or is Audi taking financial responsibility? Anyway, I personally could care less what you call v9, I would take it over this AUDI L3 without blinking. Same for the Cadillac Super cruise (whatever level they car theirs).
First L3 Self Driving Car - Audi A8 world premieres in Barcelona

 

[Az_Rael]

I did not go back and read this thread about AUDI A8 L3. BUT, is the New AUDI A8 really L3 based on the definitions in THIS Thread? If it is L3 is it because it allows the driver to do other things until the driver is needed even though this only works (I believe) in heavy traffic going < 40mph? Or is Audi taking financial responsibility? Anyway, I personally could care less what you call v9, I would take it over this AUDI L3 without blinking. Same for the Cadillac Super cruise (whatever level they car theirs).
First L3 Self Driving Car - Audi A8 world premieres in Barcelona

Eh, I think being able to read a book in slow stop and go traffic would be awesome. You could actually zone out in stop and go like a passenger. I hope Tesla works on something similar - I would consider that $3k well spent if that was all the FSD we ever got. I believe Audi is taking on financial liability for that system, although I think they are having some regulatory trouble bringing it to the US.

 

[rnortman]

Eh, I think being able to read a book in slow stop and go traffic would be awesome. You could actually zone out in stop and go like a passenger. I hope Tesla works on something similar - I would consider that $3k well spent if that was all the FSD we ever got. I believe Audi is taking on financial liability for that system, although I think they are having some regulatory trouble bringing it to the US.

The Audi system has, in addition to ultrasonics and cameras, 5 radars and a lidar, plus extra redundancy in critical systems. This gives them sensor diversity (in addition to the outright redundancy), which is the only way anybody can hope to offer L3 with today's technology. Tesla is counting on still-uninvented (or at least unproven) breakthroughs to do it without this kind of sensor suite.

 

[mongo]

The Audi system has, in addition to ultrasonics and cameras, 5 radars and a lidar, plus extra redundancy in critical systems. This gives them sensor diversity (in addition to the outright redundancy), which is the only way anybody can hope to offer L3 with today's technology. Tesla is counting on still-uninvented (or at least unproven) breakthroughs to do it without this kind of sensor suite.

My self driving system operates using only 2 cameras (and usually only one, or at least not in stereoscopic mode), so it's pretty proven to be possible.

 

[Tslrsx]

The Audi system has, in addition to ultrasonics and cameras, 5 radars and a lidar, plus extra redundancy in critical systems. This gives them sensor diversity (in addition to the outright redundancy), which is the only way anybody can hope to offer L3 with today's technology. Tesla is counting on still-uninvented (or at least unproven) breakthroughs to do it without this kind of sensor suite.

Using lidar is also unproven and Will Be for a while (just like any other system). I don't think lidar is good enough for all weather use but extra radars Will probably have a benefit in redundancy.

 

[Unpilot]

I remember leaving party years ago and the only way to get my camera's to work was to close one eye..I mean camera

 

[rnortman]

My self driving system operates using only 2 cameras (and usually only one, or at least not in stereoscopic mode), so it's pretty proven to be possible.

Only with technology that we do not yet understand. This is why I say Tesla is relying on uninvented breakthroughs.

 

[rnortman]

Using lidar is also unproven and Will Be for a while (just like any other system). I don't think lidar is good enough for all weather use but extra radars Will probably have a benefit in redundancy.

Nobody has a lidar-only system. Lidar is used in combination with radar and cameras. Sensor diversity is the key ingredient here.

 

[akraut]

Lidar is easily hamstrung by things like Rain, Fog, Steam, Smoke, etc. Hilariously, driving on city streets the lidar sees these columns of thick, unpassable objects rising from just about every manhole cover. Cameras and radar will see nothing there at all in most cases.

 

[mongo]

Only with technology that we do not yet understand. This is why I say Tesla is relying on uninvented breakthroughs.

And yet, you claim that Ultrasonic+Lidar+Radar+camera is the only hope for autonomy... Hope means it is not a proven system either...

The Audi system has, in addition to ultrasonics and cameras, 5 radars and a lidar, plus extra redundancy in critical systems. This gives them sensor diversity (in addition to the outright redundancy), which is the only way anybody can hope to offer L3 with today's technology.

With only two (one) cameras as the sensor inputs, it is proven to be possible to drive a car (given sufficient processing).
Agree to disagree?

 

[APotatoGod]

With only two (one) cameras as the sensor inputs, it is proven to be possible to drive a car (given sufficient processing).
Agree to disagree?

A deaf person with one eye is capable of driving. Thus I’d argue that it’s definitely possible with one camera (possibly more, as humans can turn their head?) to drive a car. However that doesn’t mean that that would be the best solution, nor the technologically simplest to achieve. I’d say the more sensors the better, but it should definitely be possible to control a car with only the sensors a human has.

 

[emmz0r]

A deaf person with one eye is capable of driving. Thus I’d argue that it’s definitely possible with one camera (possibly more, as humans can turn their head?) to drive a car. However that doesn’t mean that that would be the best solution, nor the technologically simplest to achieve. I’d say the more sensors the better, but it should definitely be possible to control a car with only the sensors a human has.

I don't get the people needing LIDAR. I've never seen drivers with necks sticking out of the roof with spinning heads drive around.