Autonomous Car Progress

[Tam]

Curves of death = curve where entering at the speed limit is entering too fast.
If stuff is flying around inside the car you're pushing it, and any sudden loss of traction could cause you to exit the road.

Responsible humans recognize the difficult conditions and slow down. They don't need sharp bend signs.

This thread is about the progress of autonomy.

With Tesla's slow technological progress, consumer scapegoating is a very tempting lame excuse.

The China video below is with traditional firmware (non-FSD beta) and both its map on the right and its instrument clusters clearly show that the road is curvy but the car still drove in a straight line and crashed.

Below: The driver had to take over to prevent from crashing into the iron fence. Notice the system clearly draws the red sharp 90-degree turn road shoulders. Oakland, CA area:

In slow speed, 5 different Tesla on Autopilot crashed in the same spot in Yosemite National Park in 1 month:

It's true that in "curves of death", none of the drivers died but some cars had to be towed. So, technically it's "curves of disengagement".

It's true that if drivers pay attention, the accidents wouldn't happen but we are talking about the progress of autonomy, not about the progress of drivers' skills.

 

[EVNow]

consumer scapegoating is a very tempting lame excuse.

Thats sounds exactly like what TSLAQ would say.

Its pathetic that drivers don't want to take responsibility.

We have a Toyota van - which has a lane correction system. It fails to do that correction a lot more times than a Tesla. Doesn't mean Toyota is making a "lame excuse" if a driver crashes.

 

[Tam]

...Its pathetic that drivers don't want to take responsibility...

Then open up another thread "Driver Responsibility Progress".

This thread is to hold the technology responsible and to gauge which ones are progressing better than others. Which one has progressed enough to avoid more crashes?

 

[EVNow]

Then open up another thread "Driver Responsibility Progress".

This thread is to hold the technology responsible and to gauge which ones are progressing better than others. Which one has progressed enough to avoid more crashes?

Not really. It is to figure out the progress in all dimensions.

Oh, you are welcome to start a thread on how "consumer scapegoating is a very tempting lame excuse.".

We actually have 3 dimensions
- Scenario (or feature)
- Location
- Individual (i.e. people)

Reliability is not a dimension - its a "measure". For a given combination of the above 3 dimensions we get a particular reliability (or error probability) number. Ofcourse the reliability can be calculated for a group of individuals/scenarios/locations or any combinations of them.

Waymo has high reliability for a lot of features - but only in small # of locations and only for their own cars.

ps : Elevators work in a lot of locations but in extremely limited number of scenarios

 

[mark95476]

"Curves of disengagement" doesn't even reach the threshold for FUD.

It's true that in "curves of death", none of the drivers died but some cars had to be towed. So, technically it's "curves of disengagement".

It's true that if drivers pay attention, the accidents wouldn't happen but we are talking about the progress of autonomy, not about the progress of drivers' skills.

 

[EVNow]

It's true that if drivers pay attention, the accidents wouldn't happen but we are talking about the progress of autonomy, not about the progress of drivers' skills.

So your idea is to arbitrarily spray this thread with crash images and you would call that tracking "progress" ?

 

[Tam]

So your idea is to arbitrarily spray this thread with crash images and you would call that tracking "progress" ?

Exactly! The progress for Tesla is drivers still have to be very vigilant, nowhere near Level 3.

There's another curve of death report:

FSD Beta Attempts to Kill Me; Causes Accident

Long time lurker, first time poster. I have been trying to work with Tesla to resolve this issue out of the public domain, but they have been characteristically terrible and honestly don't seem to care. Over the last 3 weeks, I have sent multiple emails, followed up via phone calls, escalated...

teslamotorsclub.com

But in my experience, I don't need any curve at all and it still happens to me. I was on a very clear nice straight road in the city with a very nice clear daylight vision and good weather, my FSD beta just suddenly jerked its steering wheel and swerved into the oncoming traffic on the opposite side that's separated by the double yellow line paint. Fortunately, unlike the driver above, I did not overcorrect the steering and it was fine for me but I got a very loud protest from my passengers.

 

[EVNow]

Exactly! The progress for Tesla is drivers still have to be very vigilant, nowhere near Level 3.

I believe we all know that. Thats not news !!!

And competitors can't even handle simple bends.

 

[Tam]

I believe we all know that. Thats not news !!!

And competitors can't even handle simple bends.

Exactly! They do not! That is why we have this thread!

Why not?

1) Ford BlueCruise refuses to have a LIDAR
2) Since Ford BlueCruise refuses to have a LIDAR, they should have followed what GM does: Have LIDAR pre-map the area first and Ford BlueCruise refuses to do LIDAR pre-mapping.

All it takes is just some homework for Ford BlueCruise.

It's homework, not news!

 

[Daniel in SD]

And competitors can't even handle simple bends.

Maybe they're all programmed to comply with EU regulations.

 

[EVNow]

1) Ford BlueCruise refuses to have a LIDAR

LOL - LiDAR to handle bends. Thats hot off the press news !

Pls send it out to Karpathy - he probably doesn't know they need LiDAR to handle simple bends.

 

[stopcrazypp]

Exactly! They do not! That is why we have this thread!

Why not?

1) Ford BlueCruise refuses to have a LIDAR
2) Since Ford BlueCruise refuses to have a LIDAR, they should have followed what GM does: Have LIDAR pre-map the area first and Ford BlueCruise refuses to do LIDAR pre-mapping.

All it takes is just some homework for Ford BlueCruise.

It's homework, not news!

Lidar is irrelevant, it's not tasked with identifying lane lines or steering the vehicle. The Ford system already whitelists areas they can use (just like Supercruise). In Munro's interview with the Ford engineer, the gist I got was it has to do with the maximum steering angle they tuned their system to use, which makes it so it simply is not able to physically make a curve at speed when the radius exceeds a certain angle (with perception being irrelevant).

This reminds me of the test years ago comparing Tesla on AP1 (using Mobileye's EyeQ3 with one camera) where Tesla would successfully navigate a curve, while the Mercedes on Drive Pilot would happily plow into the oncoming lane and even go fully into that lane.
Fatal Flaw in AP2 Design

FYI Ford uses EyeQ4 which is a more advanced chip than Tesla used on AP1.

People need to stop treating Lidar as some kind of silver bullet that would magically solve all problems with semi-autonomous/autonomous vehicles. If it was, this whole thread wouldn't have been necessary at all as years ago many players have started using it already.

 

[Tam]

LOL - LiDAR to handle bends. Thats hot off the press news !

Pls send it out to Karpathy - he probably doesn't know they need LiDAR to handle simple bends.

Like I said, as of today, to cross a double yellow line, Tesla doesn't require a curve of death: A nice clear straight road would do.

That's why with early Tesla versions, the manual says it's only for divided highways with a physical barrier.

Tesla has progressed to a degree that there's no need for a physical barrier to separate the FSD beta from the opposite traffic, a paint line would do but drivers still just need to be vigilant.

 

[stopcrazypp]

Like I said, as of today, to cross a double yellow line, Tesla doesn't require a curve of death: A nice clear straight road would do.

That's why with early Tesla versions, the manual says it's only for divided highways with a physical barrier.

Tesla has progressed to a degree that there's no need for a physical barrier to separate the FSD beta from the opposite traffic, a paint line would do but drivers still just need to be vigilant.

The things FSD Beta does has to do path planning errors (it making deliberate lane changes), nothing to do with the "curve of death" situations described above, where the car is tasked to stay in its given lane (and has no ability to make a decision to change lanes by itself), but it just fails because it was going too fast for the curve.

 

[Daniel in SD]

People need to stop treating Lidar as some kind of silver bullet that would magically solve all problems with semi-autonomous/autonomous vehicles. If it was, this whole thread wouldn't have been necessary at all as years ago many players have started using it already.

It's the silver bullet that tells you exactly where everything is (assuming it's not raining too hard? no idea what the current state of the art is). There's a reason Waymo went 6 million miles without hitting a curb (or any other stationary object). The biggest mistake people make around here is thinking that knowing exactly where everything is, what it is, and its current trajectory is the silver bullet. That's the easy part!

You could use LIDAR to tell you exactly where lane lines are. The road isn't flat so using cameras alone is reliant on NNs and/or maps.

 

[stopcrazypp]

It's the silver bullet that tells you exactly where everything is (assuming it's not raining too hard? no idea what the current state of the art is). There's a reason Waymo went 6 million miles without hitting a curb (or any other stationary object). The biggest mistake people make around here is thinking that knowing exactly where everything is, what it is, and its current trajectory is the silver bullet. That's the easy part!

You could use LIDAR to tell you exactly where lane lines are. The road isn't flat so using cameras alone is reliant on NNs and/or maps.

While some LIDAR sensors have enough resolution and provide enough data (using reflectivity) to detect lane lines, the kind of consumer LIDAR that are being used in massed produced passenger cars today (like the Valeo Scala used by Audi and no doubt would be the type Ford would use if they opted for it) simply does not have the resolution to reliably detect lane lines even if they were perfect and on a clear day with clean roads. You can see some sample outputs in Valeo's presentations:
https://www.mathworks.com/content/d...gs/point-cloud-processing-using-hdl-coder.pdf

This is a task that can be done (and had been done) however by even a relatively low resolution camera even when lane lines are faded or non-existent (that was the whole selling point of the Mobileye system). Again, perception is not the problem faced by the curve problem described above that other manufacturers still face. Tesla was able to handle that on AP1 with a single camera using the EyeQ3 chip.

 

[Daniel in SD]

While some LIDAR sensors have enough resolution and provide enough data (using reflectivity) to detect lane lines, the kind of consumer LIDAR that are being used in massed produced passenger cars today (like the Valeo Scala used by Audi and no doubt would be the type Ford would use if they opted for it) simply does not have the resolution to reliably detect lane lines even if they were perfect and on a clear day with clean roads. You can see some sample outputs in Valeo's presentations:
https://www.mathworks.com/content/d...gs/point-cloud-processing-using-hdl-coder.pdf

This is a task that can be done (and had been done) however by even a relatively low resolution camera even when lane lines are faded or non-existent (that was the whole selling point of the Mobileye system). Again, perception is not the problem faced by the curve problem described above that other manufacturers still face. Tesla was able to handle that on AP1 with a single camera using the EyeQ3 chip.

Not saying the LIDAR needs to see the lines itself. The LIDAR just tells you where the road is and therefore gives the surface to map the camera view of the lane lines onto.
I only use AP on interstates where curves can be done well over the speed limit. I guess I'll have to try it on a curvy road with hills to see how well it works. My impression from reading posts around here is that it's not all that reliable (sometimes too fast, sometimes too slow). It never seems to me that it can see more than 200ft ahead.

 

[EVNow]

I only use AP on interstates where curves can be done well over the speed limit. I guess I'll have to try it on a curvy road with hills to see how well it works. My impression from reading posts around here is that it's not all that reliable (sometimes too fast, sometimes too slow). It never seems to me that it can see more than 200ft ahead.

I’ve been using AP on curvy, hilly roads for 2+ years. Hardly ever had a problem. Not talking about hairpin bends where it apparently has issues - just normal suburban roads that are guided for 20+ mph.

 

[Bladerskb]

I believe we all know that. Thats not news !!!

And competitors can't even handle simple bends.

View attachment 742946

its funny how tesla fans always pick on bluecruise. like how do you consistently pick out one of the worse performing ADAS system out of the dozens and then defining it as "the competitors".

 

[Bladerskb]

While some LIDAR sensors have enough resolution and provide enough data (using reflectivity) to detect lane lines, the kind of consumer LIDAR that are being used in massed produced passenger cars today (like the Valeo Scala used by Audi and no doubt would be the type Ford would use if they opted for it) simply does not have the resolution to reliably detect lane lines even if they were perfect and on a clear day with clean roads. You can see some sample outputs in Valeo's presentations:
https://www.mathworks.com/content/d...gs/point-cloud-processing-using-hdl-coder.pdf

2017 called...they want their lidar back.

All cars going into production today and in the near future uses high resolution lidars, not the 4 line lidar of Scala 1. To understand the difference in resolution and spec, Luminar which will be going into several cars in 2022 has 640 lines.

Current Production Cars in 2021 with lidar:
Lucid Motors (High resolution)
Xpeng P5 (High resolution Livox lidars)
Huawei Arcfox As Hi (High resolution Huawei Lidars)

Cars releasing in 2022 with lidar (not a complete list):
Volvo (High Resolution Luminar lidar)
Nio ET7 (High resolution Innovusion lidar)
BMW IX (High resolution Innoviz lidar)

Cars releasing in 2023 with lidar (not a complete list):
GM Lyriq (High Resolution Ceptron lidar)

All the lidars are high resolution and can read lane lines, road markings, curbs, etc

Luminar Lidar output:

Innoviz Lidar output:

Innovusion Lidar output:
nio.com/cdn-static/mynio/videos/nad/nad-lidar-highway.mp4

Livox Lidar output:

the kind of consumer LIDAR that are being used in massed produced passenger cars today (like the Valeo Scala used by Audi and no doubt would be the type Ford would use if they opted for it)

Not sure how you are coming to this conclusion.