Cadillac Super Cruise vs Tesla Autopilot

[YYZ-IAD]

The following is Cadillac's brief summary of their Super Cruise feature.

What surprised me is the use of LiDAR map data. We know Caddy doesn't have LiDAR sensors -- but I wasn't aware their algorithms use LiDAR generated map data.

Q: Does Autopilot use high definition maps in its "sensor set" ??


Super Cruise

The 2021 Escalade enters the future of mobility as the first full-size SUV with Super Cruise driver assistance technology. It enables hands-free driving on more than 200,000 miles of compatible highways in the United States and Canada, using LiDAR map data, high-precision GPS, a state-of-the-art driver attention system and a network of cameras and radar sensors.

 

[diplomat33]

The following is Cadillac's brief summary of their Super Cruise feature.

What surprised me is the use of LiDAR map data. We know Caddy doesn't have LiDAR sensors -- but I wasn't aware their algorithms use LiDAR generated map data.

Q: Does Autopilot use high definition maps in its "sensor set" ??

Super Cruise

The 2021 Escalade enters the future of mobility as the first full-size SUV with Super Cruise driver assistance technology. It enables hands-free driving on more than 200,000 miles of compatible highways in the United States and Canada, using LiDAR map data, high-precision GPS, a state-of-the-art driver attention system and a network of cameras and radar sensors.

It is well known that Supercruise uses lidar map data.

As far as I know, Autopilot does not currently use HD maps but Elon did say in a recent tweet that Tesla could use "micro-maps". So Tesla might use HD maps in the future.

 

[Puddles]

What does it even mean to “use a lidar map“? A map is a semi static representation of routes, and lidar is used to visualize transient objects such as vehicles, pedestrians, bicyclists, etc.

 

[diplomat33]

What does it even mean to “use a lidar map“? A map is a semi static representation of routes, and lidar is used to visualize transient objects such as vehicles, pedestrians, bicyclists, etc.

lidar can also map lanes. So Cadillac can use lidar to map lanes very precisely. Supercruise requires that the driver center the car in the lane first before activating the system. So once you are centered in the lane, Supercruise can use that high precision map of the lanes + high precision GPS + camera vision to steer the car and make sure it stays in the right lane at all times.

 

[ewoodrick]

Cadillac is depending on high quality map data and light to moderate vision data to drive.
Tesla depends on low to mid quality map data an heavy vision to drive.

Humans depend on light quality map data and heavy vision to drive.

Which one has an issue if they change the roads?

 

[YYZ-IAD]

Which one has an issue if they change the roads?

Agree. Figuring out where the lanes are seem to be within capability of good cameras and other common sensors. As humans, we dont stare at a map to figure out where are the lane markings. That said, Waymo and other L4/5 applications do make strong use of HD maps.

 

[Puddles]

Over the last year, the 50 mile corridor I use to commute has had at least 5% under construction at some point during the entire year. Is a lidar map based system flexible enough to accommodate that dynamic environment? I’ll defer judgment right now, but the lidar mapping to me does not pass the sniff test at this point.

 

[diplomat33]

Cadillac is depending on high quality map data and light to moderate vision data to drive.
Tesla depends on low to mid quality map data an heavy vision to drive.

Humans depend on light quality map data and heavy vision to drive.

Which one has an issue if they change the roads?

Agree. Figuring out where the lanes are seem to be within capability of good cameras and other common sensors. As humans, we dont stare at a map to figure out where are the lane markings. That said, Waymo and other L4/5 applications do make strong use of HD maps.

Computer vision is really hard. We take it for granted because human vision is so well developed but it's really hard for computers. For example, if you look at raw data from Autopilot, our Teslas will be unsure about the distance of double stacked traffic lights or mistake a sign with the letter "F" for a traffic light. It makes sense to reinforce camera vision with other aids like HD maps.

HD Maps can be very useful in lots of different situations. There are plenty of cases where lane markings might be faded, missing or just outright confusing.

Take this case for example:

Can you imagine Autopilot trying to stay in the lane and ping ponging like crazy?

Or this intersection:

It might be obvious to humans but it will be confusing to computers.

or what about an intersection with faded lane markings?

Sure, you can eventually train the camera vision but it takes a lot of time and there will be cases that stump it until you train it again. So, it makes perfect sense to pre-map areas, build that HD map, to give the car the information to help it navigate in case the camera vision gets confused. HD Maps will give you much higher reliability in your lane keeping.

Again, you are not relying solely on HD maps. But you are using them as a back up to help your computer vision in case it gets confused.

 

[ewoodrick]

Computer vision is really hard. We take it for granted because human vision is so well developed but it's really hard for computers. For example, if you look at raw data from Autopilot, our Teslas will be unsure about the distance of double stacked traffic lights or mistake a sign with the letter "F" for a traffic light. It makes sense to reinforce camera vision with other aids like HD maps.

HD Maps can be very useful in lots of different situations. There are plenty of cases where lane markings might be faded, missing or just outright confusing.

Take this case for example:

Can you imagine Autopilot trying to stay in the lane and ping ponging like crazy?

Or this intersection:



It might be obvious to humans but it will be confusing to computers.

or what about an intersection with faded lane markings?

Remember the way that the Cadillac solution handles all these issues.

Sorry, does not compute.
Road not supported. PERIOD.

No surface streets for Cadillac.

 

[diplomat33]

Remember the way that the Cadillac solution handles all these issues.

Sorry, does not compute.
Road not supported. PERIOD.

No surface streets for Cadillac.

Yes, I know that. But that's my point. That's precisely why Cadillac uses lidar maps: to avoid issues where the camera vision might fail. Without a HD map, there is a chance the computer vision might fail. To avoid that, Cadillac only uses the system on those roads that have been mapped out in advance. it's designed to improve safety and reliability.

 

[ewoodrick]

Yes, I know that. But that's my point. That's precisely why Cadillac uses lidar maps: to avoid issues where the camera vision might fail. Without a HD map, there is a chance the computer vision might fail. To avoid that, Cadillac only uses the system on those roads that have been mapped out in advance. it's designed to improve safety and reliability.

In Birmingham AL, they have removed some sections of roadway bridges for construction.

Crap! the Cadillac said.

 

[Puddles]

@diplomat33 Your point about mapping things other than routes is well taken. I've been puzzling over just how a visual system will match up that left turn signal UP THERE with the (possibly misaligned) turn lane DOWN HERE, and I can see how an HD map could help with that.

It's going to be an interesting decade!

I think in the long run Elon will be right and visual processing will end up being superior. But that's going to take leaps in logic and technology, and for now I can see where the HD system could make sense.

That's it in a nutshell, isn't it? Tesla attracts dreamers, and Caddy, well, for pete's sake grandpa....

 

[ewoodrick]

@diplomat33 Your point about mapping things other than routes is well taken. I've been puzzling over just how a visual system will match up that left turn signal UP THERE with the (possibly misaligned) turn lane DOWN HERE, and I can see how an HD map could help with that.

I wonder how a vision system could match it up.

Oh wait, humans use vision systems.

 

[diplomat33]

I think in the long run Elon will be right and visual processing will end up being superior. But that's going to take leaps in logic and technology, and for now I can see where the HD system could make sense.

Like I said in the other thread, there are basically two approaches to FSD right now:

Approach #1 (Waymo, Cruise etc): Use cameras, radar, lidar and HD maps to give the car redundant reliability and excellent perception and then build the driving policy. This approach is more costly and requires pre-mapping areas in advance but it offers real autonomous driving NOW that is much safer and reliable.

Approach #2 (Tesla): Use machine learning and fleet data to develop a robust computer vision so that the car can be autonomous with just cameras. This approach is risky because right now computer vision alone is not good enough and who knows how long it will take to get there. So it will take longer and be less reliable at first. You'll have phantom braking, hitting stopped cars etc... But if you can get super computer vision, then you will have a cheap, easy to install on cars, and generalized autonomous driving system that can work everywhere. Obviously, for Tesla this would be a huge WIN if they do it.

I would point out that even if you get super computer vision, having HD maps or lidar will still make your system better. After all, even if you have excellent computer vision, having aids like HD maps or LIDAR that add an extra layer of redundancy can't be a bad thing. So it is likely that HD maps and LIDAR will still be needed anyway just to make the system even more reliable, safer and more robust. Again, you can't have too much safety. Let's pretend that you can do autonomous driving that is 3x safer than average human drivers with cameras only, but adding HD maps and LIDAR makes it 5x safer. Will you really turn down adding HD Maps and LIDAR? No. 5x safer is better than 3x safer. So if adding them makes your car safer, that will always be a good thing no matter how good your camera only system is.

 

[diplomat33]

I wonder how a vision system could match it up.

Oh wait, humans use vision systems.

Yes, but humans have an amazing "vision + brain" system. Recognizing roads, cars, signs, lane markings, pedestrians, paper bag on the road, potholes etc... comes very naturally to us because humans already have a very advanced deep neural net. Computers are not so good at it yet. I am not saying it is not possible. I am just saying it will take work. That's what Tesla is working so hard to do.

 

[ewoodrick]

Yes, but humans have an amazing "vision + brain" system. Recognizing roads, cars, signs, lane markings, pedestrians, paper bag on the road, potholes etc... comes very naturally to us because humans already have a very advanced deep neural net. Computers are not so good at it yet. I am not saying it is not possible. I am just saying it will take work. That's what Tesla is working so hard to do.

??? It takes about 16 years before you can get a license to drive. That's not exactly something that comes naturally.

 

[diplomat33]

??? It takes about 16 years before you can get a license to drive. That's not exactly something that comes naturally.

I know. I was trying to say that once you are regular driver, driving feels very easy. You don't really think about every action.

But the fact that humans wait 16 years before getting their driver's license kinda underscores the fallacy in the analogy that humans can drive with just 2 "cameras" and a brain and therefore camera only is enough to do FSD. Yes, 2 eyes and a brain is enough to drive but our brain has developed for 16 years. You can't really compare a human brain that has been trained for 16 years with computer vision that has been trained for 2-3 years. Now, I am not saying that it will take Tesla 16 years to do FSD. There is a lot that the human brain can do that is not relevant to autonomous driving. But we should be realistic too. It will definitely take time to develop "camera only" FSD. There is a lot that will need to go into the computer vision and the driving policy before it is good enough.

 

[Daniel in SD]

I wonder how a vision system could match it up.

Oh wait, humans use vision systems.

Why can't Autopilot see that this is a semi truck? Unfortunately there are things that humans cannot do yet, like make a computer vision systems reliable enough for autonomous driving.
Now there's a good chance that Autopilot calculated that there was some probability this was a semi truck and also that there was some probability it was an overpass. If your LIDAR map says that there is no overpass there then that makes it more likely to recognize it as a semi truck.

 

[Puddles]

... you can't have too much safety.

Oh sure you can. Go too far down that road and you'll end up crouching in your bunker muttering... can't sleep... clowns gonna get me...

 

[diplomat33]

Here is a slide from a presentation from about 3 years ago that compares Tesla's AP2 to Ford's Autonomous Vehicle:

We can see that if Tesla can develop excellent computer vision, then Tesla's approach could lead to full autonomous driving at a lower cost and with better EV range, two big advantages for Tesla. Of course, getting that camera vision is the hard part. Other companies that use lidar have a more expensive solution that cuts down on range but it is a faster path to FSD and achieves higher safety.

You can watch the 10 minute presentation here: