Autonomous Car Progress

[stopcrazypp]

Or the perception NN can’t tell the position of a curb 10 feet away or a bollard directly in front of the car. Because the state of the art in camera based computer vision isn’t sufficient for self driving and nobody knows when it will be.

Or they can just throw in an inexpensive parking camera system and solve at least the curb problem (and likely the bollard problem too) and give a bonus of being able to do pull in parking (instead of back-in being the only viable way or needing cars or walls on both sides and using ultrasonic sensors, also likely would improve autopark overall drastically).

For other driving, the margin of error is a lot higher so the NNs don't have to be as accurate.

Edit: depending on which option they pick for parking camera mounting, this can also solve the alleyway pull out problem discussed previously in other threads:

 

[Matias]

Because the state of the art in camera based computer vision isn’t sufficient for self driving and nobody knows when it will be.

2019

 

[Dewg]

I was just reading the NHTSA report for ADS systems released at the same time as the ADAS report that's gotten much discussion on these forums for Tesla topping the list.

I guess I wasn't surprised that Waymo topped the list, as they're in the lead for ADS systems. But I was surprised at the sheer number of companies with ADS systems. The industry definitely has some good, healthy competition.

https://www.nhtsa.gov/sites/nhtsa.gov/files/2022-06/ADS-SGO-Report-June-2022.pdf

 

[diplomat33]

Here is an interesting graph that shows ADAS based on computing and sensing. We can see that Tesla is at the bottom in terms of computing and sensing.

 

[voyager]

Request for minimum safeguards (WP editorial):

https://www.washingtonpost.com/opinions/2022/06/26/problem-with-self-driving-cars-many-dont-drive-themselves/

"The issue, it appears, may not be merely that automated systems themselves have flaws but also that drivers are relying too heavily on systems that aren’t designed to do all the work without human input. After all, when something is called “full self-driving” it’s easy to expect, consciously or subconsciously, that it will fully drive itself."

 

[Dewg]

Here is an interesting graph that shows ADAS based on computing and sensing. We can see that Tesla is at the bottom in terms of computing and sensing.

The Chinese company NIO has an impressive ADAS system:

The autonomous driving system in the NIO ET7 is called Aquila, which uses 33 sensors, including seven 8 MP high-definition cameras, four 3 MP light-sensitive surround-view cameras, a high-resolution LIDAR with a 500 m (1,640.4 ft) range, 5 millimeter-wave radars, 12 ultrasonic radars, and 2 high-precision positioning units. The Aquila autonomous driving system generates 8 gigabytes of data per second which is analyzed by an onboard computer called Adam.

 

[Dewg]

Request for minimum safeguards (WP editorial):

https://www.washingtonpost.com/opinions/2022/06/26/problem-with-self-driving-cars-many-dont-drive-themselves/

"The issue, it appears, may not be merely that automated systems themselves have flaws but also that drivers are relying too heavily on systems that aren’t designed to do all the work without human input. After all, when something is called “full self-driving” it’s easy to expect, consciously or subconsciously, that it will fully drive itself."

That's the inherent problem with all L2 systems - complacency. Doesn't matter if it's called "full self-driving" or some other industry terminology. Humans will become complacent once the car starts driving itself and they see the success rate. I'm sure everyone enables their L2 systems for the first time with a sense of trepidation, even when called Autopilot or FSD, but then learn to trust the system.

I am a very diligent driver when using AP/FSD Beta, but even I admit that I sometimes trust the system enough to let my eyes wander to some interesting feature on my left or right, and let that gaze linger a bit longer than I should. I think it's just human nature.

 

[AlanSubie4Life]

The Chinese company NIO has an impressive ADAS system:

The autonomous driving system in the NIO ET7 is called Aquila, which uses 33 sensors, including seven 8 MP high-definition cameras, four 3 MP light-sensitive surround-view cameras, a high-resolution LIDAR with a 500 m (1,640.4 ft) range, 5 millimeter-wave radars, 12 ultrasonic radars, and 2 high-precision positioning units. The Aquila autonomous driving system generates 8 gigabytes of data per second which is analyzed by an onboard computer called Adam.

Yeah if this fully autonomous driving thing actually turns out to be possible, NVIDIA will probably be pretty well positioned. I guess we’ll see! Seems like a better business prospect for them than Bitcoin, anyway.

 

[drtimhill]

Here is an interesting graph that shows ADAS based on computing and sensing. We can see that Tesla is at the bottom in terms of computing and sensing.

What was the numerical basis for both axes of this graph?

 

[diplomat33]

What was the numerical basis for both axes of this graph?

We don't know. The axes are qualitative, not quantitative.

 

[diplomat33]

What was the numerical basis for both axes of this graph?

These examples give us some idea. We can see that the cars at the top of the chart boast computers with over 1000 TOPS, compared to Tesla's 144 TOPS. Also, the cars are equipped with multiple lidar, HD radar and 8MP cameras, compared to Tesla's measly 1.2 MP cameras.

https://twitter.com/x/status/1541468208158359552

 

[Dewg]

These examples give us some idea. We can see that the cars at the top of the chart boast computers with over 1000 TOPS, compared to Tesla's 144 TOPS. Also, the cars are equipped with multiple lidar, HD radar and 8MP cameras, compared to Tesla's measly 1.2 MP cameras.

https://twitter.com/x/status/1541468208158359552

I am curious where the LIDAR is on some of these new vehicles. My understanding of how LIDAR works is very much like what we see on Waymo, and even on my home Neato Robotics vacuum cleaner, a circular device that sends light (lasers) in all directions and then picks up the bounce-backs from any direction to create a 3D model of the surrounding area.

 

[diplomat33]

I am curious where the LIDAR is on some of these new vehicles. My understanding of how LIDAR works is very much like what we see on Waymo, and even on my home Neato Robotics vacuum cleaner, a circular device that sends light (lasers) in all directions and then picks up the bounce-backs from any direction to create a 3D model of the surrounding area.

There are two kinds of lidar. There is the spinning type that Waymo uses. That is what you are thinking of. The lidar spins around, "painting" the environment with lasers in all directions. These lidars are more expensive but have the advantage of greater coverage as they can do 360 degrees. But they are bulkier too. The other type of lidar is solid state and it is basically a square that sends out lasers in a narrow beam. They are cheaper and smaller but cover a smaller angle. But they are more ideal for consumer cars since they are cheaper and can fit in the car without "ruining" the style.

The Hozon Neta S has the lidar in the front bumper. They are those squares I've highlighted.

The R7 has a forward facing lidar in that bump in the roof, right above the windshield:

 

[drtimhill]

We don't know. The axes are qualitative, not quantitative.

hmmm .. so probably have no meaning whatsoever .. the x-asis could be cost, number of sensors .. who knows what? Is the y-axis based on NN, CPU or GPU performance, or some undefined blend of each?

 

[diplomat33]

I would add that the two lidars have difference uses:

Note that the bigger, spinning lidar, we see on Waymo is placed on top and in the center of the roof. This gives the lidar a height advantage. The spinning lidar can more easily cover 360 degrees and at a pretty good distance too. These lidars are used to give the car a comprehensive "bird's eye" view of the entire scene.

The smaller solid state lidars we see on consumer cars are usually placed low in the front bumper. They are only forward facing and will have shorter range since they are close to the ground. The purpose is to provide redundancy for forward object avoidance and maintaining safe distance from a car in front. Some cars might place them in the rear bumper for something like rear object avoidance or cross traffic detection. A few cars like the R7 or Volvo, will place them at the top of the windshield to give them a bit more of a height advantage but they are still used for the same purpose.

 

[diplomat33]

hmmm .. so probably have no meaning whatsoever .. the x-asis could be cost, number of sensors .. who knows what? Is the y-axis based on NN, CPU or GPU performance, or some undefined blend of each?

See my post with the two examples. The y axis likely represents TOPS. The X axis likely represents number of sensors plus sensor performance.

 

[drtimhill]

See my post with the two examples. The y axis likely represents TOPS. The X axis likely represents number of sensors plus sensor performance.

Again, this is meaningless. "number of sensors plus sensor performance"??? hmm sounds like "Mine goes to 11" to me.

And no, I'm NOT defending Tesla, I'm merely noting that nothing useful or meaningful can be deduced from that chart.

 

[Supcom]

See my post with the two examples. The y axis likely represents TOPS. The X axis likely represents number of sensors plus sensor performance.

So, qualitative, with fuzzy quantitative when convenient.

In other words, as meaningless as a promotional video.

 

[diplomat33]

If anyone is interested, Mobileye has posted an informative marketing blog that details the different ADAS/ADS products they are doing:

Mobileye Solutions | From Driver Assistance to Self-Driving

Discover our full spectrum of solutions, from the driver-assistance systems in over 100 million vehicles today to the turnkey self-driving systems of tomorrow.

www.mobileye.com

Mobileye is calling their L4 consumer car system "Mobileye Chauffeur". It will have 6 eyeQ6 High or EyeQ Ultra Soc, 360 degrees cameras, 360 degrees radar/lidar, REM and RSS. They are targeting a cost by 2025 of under $6,000.

 

[Goose66]

And no, I'm NOT defending Tesla, I'm merely noting that nothing useful or meaningful can be deduced from that chart.

One piece of meaningful information reflected in this chart is that apparently Tesla is alone in their thinking that fewer sensors are better. Will Tesla end up being right? Maybe, but with the bulk of the "leaders" in the field going the opposite way, just based on basic mathematics alone I wouldn't bet on it.