Autonomous Car Progress

[stopcrazypp]

Yup, I found it interesting that NBC asked for the video footage they referenced and were told it's "proprietary material". IMO that just seems shady to me, if you have video that proves your side of a story, why turn it into a "yeah, well totally believe us" claim unless the video, again, makes Cruise look bad.

Additionally, there are multiple claims in the linked doc for this one incident. I don't see how Cruise can claim their cars never stalled, when it was a long enough stall to apparently go from arrival to the scene, to them trying to leave. Even going so far as SFPD trying to manually move the cars, and being unable to.



The linked doc has a ton of interesting commentary around some stalls as well.



They also still have issues running over fire hoses apparently.

The article questioned if the gap left after the first car moved was wide enough to fit an ambulance, so perhaps the video may still make the Cruise look bad (also would show there was still a stuck vehicle). Releasing the video kind of disarms Cruise's ability to spin, and may set future expectations to show video footage, so I can see why they refuse to do so.

From third party footage of other incidents, even when the incidents don't end up seriously harming anyone and may technically be legal, it still looks very bad on Cruise.

It seems like a lot of the incidents can be relatively quickly resolved if the emergency responders had the ability to do a manual take over which it seems from the report neither Cruise nor Waymo provide. From one of the incident reports manual take overs seem to be routinely done in normal non-AV vehicles when the original driver gets incapacitated.

 

[ThomasD]

In some areas that AV would be pushed out of the way by either the Firetruck, Ambulance or some other vehicle.

 

[flutas]

In some areas that AV would be pushed out of the way by either the Firetruck, Ambulance or some other vehicle.

Normally I'm not one for the "use of force" way, but imo if it's a stalled car with nobody in it, they should push it out of the way and send them to bill for any damage to the ambulance / truck / whatever.

Basically move the "f*** around and find out" scale into the world of AVs stalling, I bet Cruise / others would be very quick to fix the stalls if this policy was implemented.

 

[hooty]

It would be very interesting to see if there is any correlation between certain demographics & opposition/support for AVs.
I would include age, education level, educational field, income category, IQ, citizen/immigrant, voting preference, gender, etc

 

[diplomat33]

At IAA 2023, DeepRoute.AI demoed their Driver 3.0 autonomous driving that does not use HD maps.

https://twitter.com/x/status/1698735560875983266

If you can overlook the annoying overexcited commentary, the driving scenarios is dense urban driving is pretty good.

 

[diplomat33]

Top 10 examples of NZP (Mobileye Supervision) handling difficult cases on highway from stopped car on the highways, other cars cutting in, etc...

Sina Visitor System

 

[SidetrackedSue]

Top 10 examples of NZP (Mobileye Supervision) handling difficult cases on highway from stopped car on the highways, other cars cutting in, etc...

Sina Visitor System

Link doesn't work for me. Takes me to a landing page in Chinese and no signs of a NZP demo in the thumbnails I glanced at.

 

[diplomat33]

Link doesn't work for me. Takes me to a landing page in Chinese and no signs of a NZP demo in the thumbnails I glanced at.

Weird. It works for me.

See if this link works: Sina Visitor System

Here is a summary what the video shows:
10. Bus in right lane encroaches in car's path. NZP maneuvers to the left to avoid a collision.
9. In tight bend, NZP moves over to the left lane, when cones block off right lane.
8. NZP in middle lane of 3 lane road, in traffic jam, cuts in smoothly in gap in stopped traffic.
7. NZP moves over to left lane when it sees arrows and concrete barrier in right lane.
6. Lead car in off ramp, brakes hard due to side barrier, NZP slows and then avoids side barrier.
5. At night on curvy over pass, there is stopped car in right lane, NZP changes lanes to go around it.
4. Dense traffic going around 2 car accident, NZP squeezes into right lane to go around stopped cars.
3. Lead car suddenly swerves, revealing stopped car in the road. Zeekr brakes hard and avoids collision.
2. At night, Zeekr maneuvers around stopped car in the middle of road.
1. At night, Zeekr maneuvers around a stopped car with emergency lights on.

 

[diplomat33]

https://twitter.com/x/status/1698998579245760551

 

[diplomat33]

Link doesn't work for me. Takes me to a landing page in Chinese and no signs of a NZP demo in the thumbnails I glanced at.

The video in this tweet should work.

https://twitter.com/x/status/1699014677848572142

 

[diplomat33]

https://twitter.com/x/status/1699026685331652909

 

[flutas]

The video in this tweet should work.

https://twitter.com/x/status/1699014677848572142

Honestly, feels like a clone of FSD Beta from the video, but worse in some spots and better in others.

I wasn't that impressed with any of the cases, but I'll break down why.

10: the bus was honestly bad, the car IMO should have hit the brakes while swerving, not maintaining speed that would impact the bus if the bus doesn't stop changing lanes.

9: Detecting cones in the road and changing lanes is ground breaking?

8: Missed the right time to get over, so it cuts in front of a bus, while blocking the lane of travel it was in for an unknown amount of time (gotta speed up that part of the video!)

7: It's pulling a classic FSD no confidence move here, where it just randomly nudges forward until it gets the confidence to go. Meanwhile it's cutting in to a high speed lane of travel next to it going ~12mph, while the driver is focused on making sure it doesn't hit the barrier and not checking that the lane is empty.

6: Lane keeping...is revolutionary!

5: Avoiding a stopped car by changing lanes, again revolutionary!

4: Again, classic FSD indecisiveness shown by the car swerving the tentacle left and right.

3: AEB and a safe following distance...is revolutionary!

2: Passes by a car stopping on the road by about 2cm at a high rate of speed. (Speedo is conveniently hidden in this one.)

1: (copy from #2 here) Add in to the fact from the car driving, to almost running over their emergency triangle the car only slows by 10km/h. It isn't even fully out of the same lane when it passes them.

Overall I don't see how any of this is "raising the bar" it has a lot of the same issues that FSD Beta does currently as well. Definitely not at the level of Waymo or Cruise, and I'd say probably either equivalent or worse than FSD currently.

 

[Bladerskb]

Honestly, feels like a clone of FSD Beta from the video, but worse in some spots and better in others.

I wasn't that impressed with any of the cases, but I'll break down why.

10: the bus was honestly bad, the car IMO should have hit the brakes while swerving, not maintaining speed that would impact the bus if the bus doesn't stop changing lanes.

9: Detecting cones in the road and changing lanes is ground breaking?

8: Missed the right time to get over, so it cuts in front of a bus, while blocking the lane of travel it was in for an unknown amount of time (gotta speed up that part of the video!)

7: It's pulling a classic FSD no confidence move here, where it just randomly nudges forward until it gets the confidence to go. Meanwhile it's cutting in to a high speed lane of travel next to it going ~12mph, while the driver is focused on making sure it doesn't hit the barrier and not checking that the lane is empty.

6: Lane keeping...is revolutionary!

5: Avoiding a stopped car by changing lanes, again revolutionary!

4: Again, classic FSD indecisiveness shown by the car swerving the tentacle left and right.

3: AEB and a safe following distance...is revolutionary!

2: Passes by a car stopping on the road by about 2cm at a high rate of speed. (Speedo is conveniently hidden in this one.)

1: (copy from #2 here) Add in to the fact from the car driving, to almost running over their emergency triangle the car only slows by 10km/h. It isn't even fully out of the same lane when it passes them.

Overall I don't see how any of this is "raising the bar" it has a lot of the same issues that FSD Beta does currently as well. Definitely not at the level of Waymo or Cruise, and I'd say probably either equivalent or worse than FSD currently.

While I agree with some of your points. I would like to point out that the current Tesla FSD software in China does almost none of these things. And the only other ADAS software that can do majority of these things that I am aware are Huawei NCA & Xpeng XNGP.

So while they are not claiming this is revolutionary although they use the word "stunning". If Tesla did any of these things. All the Tesla fans would be praising it as "revolutionary" (and they have claimed revolutionary for way less). So we have to keep that in mind.

 

[diplomat33]

Honestly, feels like a clone of FSD Beta from the video, but worse in some spots and better in others.

I wasn't that impressed with any of the cases, but I'll break down why.

10: the bus was honestly bad, the car IMO should have hit the brakes while swerving, not maintaining speed that would impact the bus if the bus doesn't stop changing lanes.

9: Detecting cones in the road and changing lanes is ground breaking?

8: Missed the right time to get over, so it cuts in front of a bus, while blocking the lane of travel it was in for an unknown amount of time (gotta speed up that part of the video!)

7: It's pulling a classic FSD no confidence move here, where it just randomly nudges forward until it gets the confidence to go. Meanwhile it's cutting in to a high speed lane of travel next to it going ~12mph, while the driver is focused on making sure it doesn't hit the barrier and not checking that the lane is empty.

6: Lane keeping...is revolutionary!

5: Avoiding a stopped car by changing lanes, again revolutionary!

4: Again, classic FSD indecisiveness shown by the car swerving the tentacle left and right.

3: AEB and a safe following distance...is revolutionary!

2: Passes by a car stopping on the road by about 2cm at a high rate of speed. (Speedo is conveniently hidden in this one.)

1: (copy from #2 here) Add in to the fact from the car driving, to almost running over their emergency triangle the car only slows by 10km/h. It isn't even fully out of the same lane when it passes them.

Overall I don't see how any of this is "raising the bar" it has a lot of the same issues that FSD Beta does currently as well. Definitely not at the level of Waymo or Cruise, and I'd say probably either equivalent or worse than FSD currently.

The title "stunning scenarios" is hyperbole. A lot of it is not revolutionary. You correctly point out where it could have been better. But the key is that the AV does it and does it safely. Remember that Tesla had several high profile accidents a couple years back where regular AP crashed into stopped cars and dividers. it is not a given that a driver assist can safely handle these cases. So it is not nothing to show that SuperVision can be safe in these scenarios.

NZP is Zeekr's version of Tesla's NOA. So it will have many of the same features. And remember that Mobileye SuperVision is designed to be a supervised vision-based point-to-point driving system. So it falls into the same category as FSD Beta. So it is not surprising that it would behave a lot like a FSD Beta clone since it is the same type of product.

I think there are definitely challenges with deploying a safe and reliable supervised vision-based "FSD" system. So all these systems will have weaknesses and strengths. But I don't think we can make a definitive comparison just based on these 10 clips. Luckily, with NZP now going out to 110,000 Zeekr owners, we are getting a lot more videos from owners to make a better judgment. I do think the real challenge will be city driving. We know FSD beta handles highway driving very well. It's solving city driving that is the real challenge. That will be hard for both FSD beta and Mobileye Supervision.

 

[flutas]

While I agree with some of your points. I would like to point out that the current Tesla FSD software in China does almost none of these things. And the only other ADAS software that can do majority of these things that I am aware are Huawei NCA & Xpeng XNGP.

So while they are not claiming this is revolutionary although they use the word "stunning". If Tesla did any of these things. All the Tesla fans would be praising it as "revolutionary" (and they have claimed revolutionary for way less). So we have to keep that in mind.

Also have to keep in mind Xpeng was publicly accused of stealing Tesla's source code, and the engineer involved settled a lawsuit with and had to pay Tesla an undisclosed sum. Tesla settles lawsuit against engineer who it claims stole Autopilot source code for Chinese competitor

You are right though, they never claimed "revolutionary" that was my bad for rolling my eyes at these demonstrations being declared "stunning" that I forgot what they actually claimed because their examples were so...common place. Also, re: "Tesla FSD software in China does almost none of these things". Tesla doesn't have a FSD software that is available to demonstrate in China. Will they ever? Probably. Will it be better than this? Who knows, not me but I will say probably simply because these cases demonstrated seem...basic in comparison. I'm sure someone will beat Tesla with a better system, but I don't think it's MobileEye, at least not this one.

The title "stunning scenarios" is hyperbole. A lot of it is not revolutionary. You correctly point out where it could have been better. But the key is that the AV does it and does it safely. Remember that Tesla had several high profile accidents a couple years back where regular AP crashed into stopped cars and dividers. it is not a given that a driver assist can safely handle these cases. So it is not nothing to show that SuperVision can be safe in these scenarios.

"the AV does it and does it safely"

ah yes, swerving into a lane with a car passing you at 70kmph difference, while the driver is not watching that lane, and instead watching to make sure the car doesn't hit a stationary barrier is "safe"

blocking a flowing lane of traffic so that you don't miss your exit (creating stopped traffic for others to avoid) is "safe"

accelerating at a bus that is coming into your lane (enough that the passenger starts screaming) is "safe"

The car may not have hit anything, but I don't know if I'd say it was "safe" about doing so.

But I don't think we can make a definitive comparison just based on these 10 clips.

I was making that comparison since they lauded these 10 clips as "stunning scenarios" and "raising the bar", while you referenced them as "handling difficult cases on highway." These clips feel like common occurrences with Tesla, Waymo, Cruise, BlueCruise and other ADAS systems already on the market. I fail to see how this "raises the bar" in comparison, especially if these are the 10 clips they chose to showcase the system.

 

[diplomat33]

Some tidbits on DeepRouteAI's hardware.

https://twitter.com/x/status/1699141243970658381

 

[hooty]

Also have to keep in mind Xpeng was publicly accused of stealing Tesla's source code, and the engineer involved settled a lawsuit with and had to pay Tesla an undisclosed sum. Tesla settles lawsuit against engineer who it claims stole Autopilot source code for Chinese competitor

Tesla did not pursue a case against XPeng & XPeng stated they had no use for the code.
Makes sense since XPeng use LiDAR, so it’s all moot.

 

[SidetrackedSue]

The video in this tweet should work.

Thanks, it did! Wow, especially the braking for the stopped car that suddenly was revealed by the swerving car.

 

[stopcrazypp]

Tesla did not pursue a case against XPeng & XPeng stated they had no use for the code.
Makes sense since XPeng use LiDAR, so it’s all moot.

XPeng still uses cameras, so I disagree they had no use for the code.

The case went nowhere because Tesla didn't actually sue XPeng directly and thus there was a limit to what they could do in discovery, with the most being a third party source code comparison that turned up negative, at which point Tesla settled with the employee (with employee only paying an undisclosed fine for violating employee conduct by downloading the source code).

XPeng UI and car design is pretty clearly "inspired" by Tesla's (especially if you look at their earliest vehicles), but those elements are not legally enforceable as long as they make minor changes such that it's not a 100% exact knockoff (this game Chinese automakers have been playing for a long time in making knockoff products, so they know how to avoid legal trouble).

 

[Bladerskb]

Also have to keep in mind Xpeng was publicly accused of stealing Tesla's source code, and the engineer involved settled a lawsuit with and had to pay Tesla an undisclosed sum. Tesla settles lawsuit against engineer who it claims stole Autopilot source code for Chinese competitor

XPeng still uses cameras, so I disagree they had no use for the code.

The case went nowhere because Tesla didn't actually sue XPeng directly and thus there was a limit to what they could do in discovery, with the most being a third party source code comparison that turned up negative, at which point Tesla settled with the employee (with employee only paying an undisclosed fine for violating employee conduct by downloading the source code).

XPeng UI and car design is pretty clearly "inspired" by Tesla's (especially if you look at their earliest vehicles), but those elements are not legally enforceable as long as they make minor changes such that it's not a 100% exact knockoff (this game Chinese automakers have been playing for a long time in making knockoff products, so they know how to avoid legal trouble).

Do people actually think there are ancient AP 2018 code being used by Xpeng's SOTA system right now? Seriously? The Tesla software in 2018 doesn't do 1% of what Xpeng software does today.

Its like accusing someone who comes out today with a legit IPhone 15 competitor for stealing the source code from the original IPhone.
Or me getting an F on a test and accusing someone who got a A for copying me.

The Tesla 2018 code base and Xpeng's current system codebase are literally worlds apart. I hate that this keeps getting brought up. 5 years from now people will still be bringing this up as cope. As though Xpeng is using code from 10 years ago.