Autonomous Car Progress

[Dewg]

They are... unlike you and Tesla fans, we can approach things critically and give out constructive criticism like JJRick did and not be hanged by an angry mob. It actually proves that real ADAS/AV lovers are NOT biased towards a particular company. They can logically deduce who is ahead aswell. IF another company shows up and deploys something better, we will logically assess and crown that company.

For Tesla fans like you, Tesla is 10 years ahead. No matter what.
Tesla FSD can plow through a crowd of children and kill dozens and still be considered 10 years ahead.
Nothing will alter your mind.
That's not logic or reason.
That's called fanaticism.

I consider myself a Tesla fan, but not a fanatic. I acknowledge problems the car has, but I'm also totally impressed with the features and progress. What other car can I get behind the wheel and have it drive me around on city streets? It's definitely not perfect, but still amazes me every time I have a 0 intervention drive.

Waymo also blows me away. Watching it in person while visiting SF recently was fun. Lots of them driving around downtown - my favorite was a Waymo, then a Cruise a few cars back, then several cars behind, a Zoox. Exciting times we're in.

 

[Supcom]

Here we see Cruise phantom brake hard in the middle of an empty road at night. Maybe the trailer parked on the side freaked it out? Really odd.

That was hard to watch and I felt bad for Cruise having such a poor performance.

The routing looked really convoluted. Perhaps that's to stay on 25 mph roads? Swerving around as Tesla owners would say, "like a 15 year old." And professional grade phantom braking.

Really surprising considering the streets are so wide and simple compared to San Francisco.

 

[Bitdepth]

Is that how it is in general (including in SF) or is it from a result of overtuning to SF (in which case their bragging about only taking 90 days to launch in AZ is over exaggerated)? I don't recall similar complaints about SF, but then that may be because there haven't been the same type of review posted.

I do think that might be what it is, it should be quickly improved after they add more data from AZ to their training set. I wouldn't say it is overrated, it shows how quick they can go from not having a single presence in a region to setting up infrastructure and actually start driving without any safety driver.

You can contrast that to Waymo who takes a lot more time and tests with safety driver before they allow general public to use the service. Though they've been steadily reducing the time from testing to allowing people to use it.

 

[diplomat33]

I do think that might be what it is, it should be quickly improved after they add more data from AZ to their training set. I wouldn't say it is overrated, it shows how quick they can go from not having a single presence in a region to setting up infrastructure and actually start driving without any safety driver.

You can contrast that to Waymo who takes a lot more time and tests with safety driver before they allow general public to use the service. Though they've been steadily reducing the time from testing to allowing people to use it.

I know we debate which approach is best but I suspect both approaches probably work as long as you do it right. The "scale first then solve" approach works if you are able to solve problems quick enough and keep your system safe enough. So Cruise's challenge is to quickly improve their AV in the new areas while not causing accidents. The "solve first and then scale" approach also works if you are able to scale quick enough after solving. So Waymo's challenge is to deploy their AVs in new areas quick enough.

 

[eli_]

A drive to Taco Bell

Delivering on our first Twitter promise!

blog.comma.ai

Comma posted their hands-free drive to Taco Bell on a comma three.

 

[Dewg]

A drive to Taco Bell

Delivering on our first Twitter promise!

blog.comma.ai

Comma posted their hands-free drive to Taco Bell on a comma three.

Did you notice the GPS was off through much of the drive? It would snap back to the road occasionally, but many times was showing the car off the road.

 

[stopcrazypp]

There are a couple long unedited Cruise videos in SF that you can find on reddit but for the most part, almost all the videos we see from Cruise in SF are highly edited. They are usually just 1 mn "tick tock" style videos of someone going "OMG, the car has no driver! I am living in the future". And of course, Cruise does a great PR job of retweeting these videos to promote that they are deploying true driverless cars. However, JJ always does a great job with these reviews because, like with his Waymo videos, he will share longer unedited videos that show the good and the bad. He does much more raw and comprehensive reviews of the tech, wheter it is Waymo or Cruise.

It could be that Cruise actually is that bad in SF and they just do a good PR job of hiding it. And it could be that Cruise limits driverless at night precisely to make the ODD easier because the tech is not ready for more complicated day driving. This JJ video and the reports of Cruise stalling and blocking traffic seem to support that hypothesis.

It could also be that Cruise is super overfit to SF and so the tech works better in SF than in other places. The fact that they have spent 9 years developing autonomous driving only in SF and only recently expanded their autonomous driving outside of SF would seem to support that hypothesis.

It could also be that Cruise rushed deployment to Phoenix and Austin because they wanted to keep their timeline of deploying in 90 days. Cruise could be under a lot of pressure from GM to show that they can scale. All that pressure might have led Cruise to cut corners and maybe not test as much as they should or maybe not map certain areas. This could also explain the poor performance.

Personally, I am really surprised and kinda shocked. Driving empty residential streets is very easy. I expectected better from Cruise which has been working on AVs for 9 years.

Great points, it could be a combination of all the factors you pointed out. I think all the AV companies are under pressure to deliver something, given the economic downturn likely have investors and management questioning things if there is no progress shown in terms of launches.

 

[eli_]

Did you notice the GPS was off through much of the drive? It would snap back to the road occasionally, but many times was showing the car off the road.

Dead reckoning maybe?

The full logs of the drive is up here, with 1x speed camera captures: connect

 

[Doggydogworld]

I do think that might be what it is, it should be quickly improved after they add more data from AZ to their training set. I wouldn't say it is overrated, it shows how quick they can go from not having a single presence in a region to setting up infrastructure and actually start driving without any safety driver.

Cruise has tested in Phoenix for years. Vogt used to throw shade on Waymo's Chandler service by comparing Cruise stats in AZ (easy) vs. SF (super hard).

A couple of (former?) Cruise AZ safety drivers have posted on Reddit, etc. I think one said Cruise started AZ testing six years ago. He didn't think they were ready to deploy yet.

I don't know why they seem to do so much better in San Francisco. It is generally a tougher environment. Cruise has more experience in SF, though, plus their engineering is local and probably much more focused on SF-specific issues.

 

[Dewg]

Dead reckoning maybe?

The full logs of the drive is up here, with 1x speed camera captures: connect

I thought that, but dead reckoning would take into account position of the wheels and rotation speed to calculate position if GPS were to drop. In the YT and the full capture site, the car is moving straight with no turns, so not sure where it suddenly went left off the map in positioning. Could have been interference with GPS, or just a positional bug in the mapping data. Interesting to see that with GPS off by a decent amount, the drive seemed unaffected.

 

[stopcrazypp]

I do think that might be what it is, it should be quickly improved after they add more data from AZ to their training set. I wouldn't say it is overrated, it shows how quick they can go from not having a single presence in a region to setting up infrastructure and actually start driving without any safety driver.

You can contrast that to Waymo who takes a lot more time and tests with safety driver before they allow general public to use the service. Though they've been steadily reducing the time from testing to allowing people to use it.

It seems the two companies simply have different definitions of ready to deploy. Some of the moves the Cruise car is making (random hard braking for no apparent reason), when it happens on AP or FSD, forum goers here would scream about it being dangerous and not being fit for the road.

 

[nativewolf]

It seems the two companies simply have different definitions of ready to deploy. Some of the moves the Cruise car is making (random hard braking for no apparent reason), when it happens on AP or FSD, forum goers here would scream about it being dangerous and not being fit for the road.

True

 

[Supcom]

It seems the two companies simply have different definitions of ready to deploy. Some of the moves the Cruise car is making (random hard braking for no apparent reason), when it happens on AP or FSD, forum goers here would scream about it being dangerous and not being fit for the road.

It did, at least, respond nicely to all of the speed bumps.

 

[kabin]

The full logs of the drive is up here, with 1x speed camera captures: connect

Completing the drive is a feather in the cap but the x1 video is an eye opener. The drive very slow and overly cautious. I don't think it ever passed a vehicle on the roadway. It drove over the median line during the initial I15 South turn off and then didn't turn sharp enough moments needing to drive into the left lane. And again on the Market St off ramp it struggles making the turn and needs to brake almost to a stop.

Wonder how many times they tried before they had a drive video nearly worth uploading.

 

[diplomat33]

CNBC’s Michael Wayland recently tested ADAS from Tesla, GM and Ford:

I drove hundreds of miles 'hands-free' in GM, Ford and Tesla cars – here's how it went

CNBC's Michael Wayland recently tested "hands-free" systems from Tesla, GM and Ford. Theirs are among the most readily available and dynamic on the market.

www.cnbc.com

Overall, I think it is pretty fair. I don't agree that it is hard to know if FSD is on or not. The blue wheel icon on the screen is easy to see in your peripheral vision so you don't need to take your eyes off the road. Also, there is a very characteristic chime sound when the system turns on or off. So it is not an issue for me.

 

[Dewg]

CNBC’s Michael Wayland recently tested ADAS from Tesla, GM and Ford:

I drove hundreds of miles 'hands-free' in GM, Ford and Tesla cars – here's how it went

CNBC's Michael Wayland recently tested "hands-free" systems from Tesla, GM and Ford. Theirs are among the most readily available and dynamic on the market.

www.cnbc.com

My general takeaway from that article is that, when comparing highway/freeway driving, Ford and GM had some issues/disengagements, but were totally hands-free. Tesla was impressive and worked very well, but had wheel nags. Then the article talks about FSD Beta on city streets and the struggles 69.3.1 had navigating those city streets. The article does show that both Ford and GM don't have city streets. One small detail that was inaccurate was that the only indication that Tesla has AP/NoA/FSD Beta engaged is a little blue wheel icon, which was frustrating the author. The author doesn't indicate that the visualizations also change (AP has dual blue lines surrounding the lane the Tesla is in, and NoA/FSD Beta have a pretty thick blue line indicating where the ADAS system is going to navigate the car. Those blue lines don't exist when driving manually.

Otherwise, an interesting article to see how Ford and GM are coming along with their highway/freeway AP solution.

 

[Supcom]

CNBC’s Michael Wayland recently tested ADAS from Tesla, GM and Ford:

I drove hundreds of miles 'hands-free' in GM, Ford and Tesla cars – here's how it went

CNBC's Michael Wayland recently tested "hands-free" systems from Tesla, GM and Ford. Theirs are among the most readily available and dynamic on the market.

www.cnbc.com

Overall, I think it is pretty fair. I don't agree that it is hard to know if FSD is on or not. The blue wheel icon on the screen is easy to see in your peripheral vision so you don't need to take your eyes off the road. Also, there is a very characteristic chime sound when the system turns on or off. So it is not an issue for me.

I got the impression that the author took only a single ride in a Tesla before writing the article, despite the fact that there are many many time more cars with AP than GM and Fords combined. How could you call GM's Supercruise the 'front runner' when it disengages every 30 minutes?

Fair enough pointing out FSD beta errors, but the author did not inform readers that FSD beta was test software available only by specific request.

 

[Bladerskb]

So @Bladerskb do you have any thoughts on the JJ video in Cruise? How do you explain the issues we see in the video?

I posted this on reddit, but i will repost it here...

While this isn't an impressive ride comfort wise, some on Reddit were saying this "isn't the same driver", or that "it must be bad in SF".

But watching it i could tell that the "mistakes" its making actually prove its the same driver.

Its getting too close to the parked cars because in SF, their roads are narrow so it's used to getting close to parked cars.

Its cutting back to thread the needle after there's a gap in-between parked cars is also a side-effect of SF. Because in SF the roads are narrow so it has to cut back in most times after passing parked cars.

The planner is obviously at fault here. It just needs to be optimized a-bit more to make good use of the fact that the roads are WIDE and not narrow. So it doesn't need to do what it does in SF.

The same is true with it pointing to the right at traffic light/stop signs when it wants to go straight. Or winding left in order to turn right (another side-effect of SF driving). As we know, SF roads are narrow, steep, off-centered and twisted.

The real problems I see is the curvature in which it takes some of these turns, the absolute bonkers routing it seems and the apparent phantom brakes.

All things which I'm certain it will improve on in the coming months. The phantom brakes overall will be *eliminated by the Origin kinda how Waymo's 5th gen *eliminated them almost completely.

All things considered. Its a car with no human driver even though the ride wasn't 100% smooth,

 

[diplomat33]

I posted this on reddit, but i will repost it here...

While this isn't an impressive ride comfort wise, some on Reddit were saying this "isn't the same driver", or that "it must be bad in SF".

But watching it i could tell that the "mistakes" its making actually prove its the same driver.

Its getting too close to the parked cars because in SF, their roads are narrow so it's used to getting close to parked cars.

Its cutting back to thread the needle after there's a gap in-between parked cars is also a side-effect of SF. Because in SF the roads are narrow so it has to cut back in most times after passing parked cars.

The planner is obviously at fault here. It just needs to be optimized a-bit more to make good use of the fact that the roads are WIDE and not narrow. So it doesn't need to do what it does in SF.

The same is true with it pointing to the right at traffic light/stop signs when it wants to go straight. Or winding left in order to turn right (another side-effect of SF driving). As we know, SF roads are narrow, steep, off-centered and twisted.

The real problems I see is the curvature in which it takes some of these turns, the absolute bonkers routing it seems and the apparent phantom brakes.

All things which I'm certain it will improve on in the coming months. The phantom brakes overall will be *eliminated by the Origin kinda how Waymo's 5th gen *eliminated them almost completely.

All things considered. Its a car with no human driver even though the ride wasn't 100% smooth,

Thanks. Good insights. You mention a few times that the issues are likely caused by Cruise driving in SF. It sounds like you do think a lot of the issues are caused by Cruise overfitting to SF. Basically, Cruise has done most of their driving in SF so their autonomous driving is overfit for SF. The good news, as you point out, is that as Cruise does expand to other places like Phoenix and Austin, they will be able to retrain their autonomous driving to be more generalized.

 

[Knightshade]

My general takeaway from that article is that, when comparing highway/freeway driving, Ford and GM had some issues/disengagements, but were totally hands-free. Tesla was impressive and worked very well, but had wheel nags. Then the article talks about FSD Beta on city streets and the struggles 69.3.1 had navigating those city streets. The article does show that both Ford and GM don't have city streets. One small detail that was inaccurate was that the only indication that Tesla has AP/NoA/FSD Beta engaged is a little blue wheel icon, which was frustrating the author. The author doesn't indicate that the visualizations also change (AP has dual blue lines surrounding the lane the Tesla is in, and NoA/FSD Beta have a pretty thick blue line indicating where the ADAS system is going to navigate the car. Those blue lines don't exist when driving manually.

Otherwise, an interesting article to see how Ford and GM are coming along with their highway/freeway AP solution.

There's also audio indiications when AP is engaged or disengaged as well... one of several things that they found annoying that if you owned/drove the car for a while you'd realize aren't actual issues.

That said the only really interesting thing in there to me was just how bad Fords system still is... It was well over a year ago Sandy Munroe did a video riding in a Ford Blues Clues highway ADAS car and noted how it was incapable of handling anything more than the mildest curves on highways and would disengage often due to highway curves.... and it appears they still haven't done anything to address that.... an issue even GM was doing well on years prior and Tesla has had largely solved even longer.