Autonomous Car Progress

[willow_hiller]

Is Cruise now at the point where pulling safety drivers actually improves safety? I don't know, but it's possible. Safety drivers could certainly resolve stalls faster, which would be more convenient for others.

Maybe we should rename them convenience drivers. Or PR improvement drivers.

I think a good compromise would be something like a "Backup Driver." Where you treat it as a Level 3 system and the driver is instructed not to take control unless directed to by the vehicle.

But other than a small reduction in staffing costs (and the aforementioned PR), I can't imagine why Cruise wouldn't opt for something like it. At this point, their decision not to have a backup driver to move stopped vehicles is hurting their PR. Maybe they're hoping people will remember their jump to Level 4, and forget about the stopped vehicles over time.

 

[diplomat33]

Maybe we should rename them convenience drivers. Or PR improvement drivers.

"PR improvement driver" would be silly. But I get your point.

I think Waymo calls them "autonomous specialists". That's obviously a generic job title that does not say much. Maybe "stall prevention specialist"? LOL.

At this point, their decision not to have a backup driver to move stopped vehicles is hurting their PR. Maybe they're hoping people will remember their jump to Level 4, and forget about the stopped vehicles over time.

I feel like Cruise is taking a bit of a "FSD beta" approach. Basically, deploy first, fix issues as you go with OTA updates, and hope people forget the issues over time. and they might be correct. If say a year from now, they have 1000s of Origin vehicles doing ridesharing and the stalls are solved, will people still care about problematic stalls that happened a year ago? Probably not. Part of it might be that Cruise is likely under a lot of pressure by GM to show results. So to speed things up, they figure they can do driverless now, solve problems as they go, and get to mass deployment of the Origin sooner which is what GM wants.

 

[diplomat33]

8x speed clip of Wayve driving autonomously through busy London with vision-only end-to-end NN (ie one NN that takes vision in and outputs controls):

https://twitter.com/x/status/1669374996580298757

It does not mean that they have solved FSD. But it is a still an interesting demo of what end-to-end can do IMO. I can see why people see E2E as a promising approach.

 

[Bitdepth]

8x speed clip of Wayve driving autonomously through busy London with vision-only end-to-end NN (ie one NN that takes vision in and outputs controls):

It does not mean that they have solved FSD. But it is a still an interesting demo of what end-to-end can do IMO. I can see why people see E2E as a promising approach.

It is always amazing to see the different approaches used by different companies. Would be interesting to see how far they are able to push their camera and radar only E2E system to remove the safety personnel with current technology.

 

[Bitdepth]

At some point the "safety" driver reduces safety. A few years ago I went through every reported accident involving Waymo over a 3-6 month period. The only ones where Waymo was arguably at fault were caused by safety drivers. Sometimes the car was in manual mode the whole way. Sometime the safety driver intervened and caused a wreck. For example, one car veered to the left in slowish, multi-lane traffic. The safety driver intervened, steering the car back to the right and directly into the path of a lane-splitting motorcyclist who hit the right rear corner bumper.

Most memorable was the safety driver who accidentally disengaged the autonomous system while asleep on the highway. Fortunately he just ended up in the ditch and walked away.

The same thing happened when Waymo first moved into downtown SF. One driver felt the car was waiting too long at a green light, intervened and hit a skateboarder. All the "injury accidents" SF Transit called out in their protest letter to CPUC involved safety drivers.

Is Cruise now at the point where pulling safety drivers actually improves safety? I don't know, but it's possible. Safety drivers could certainly resolve stalls faster, which would be more convenient for others.

Maybe we should rename them convenience drivers. Or PR improvement drivers.

Said everything i was thinking.

 

[diplomat33]

It is always amazing to see the different approaches used by different companies. Would be interesting to see how far they are able to push their camera and radar only E2E system to remove the safety personnel with current technology.

I welcome different approaches. As I see it, the more approaches, the quicker we will solve autonomous driving and scale everywhere. And different approaches will have different pros and cons which can offer different value to consumers. So for example, maybe someday Wayve is able to offer supervised FSD everywhere on consumer cars. And since it is vision end-to-end, it is a a very cheap package that can easily go into affordable mass market cars. That could be very beneficial to a lot of consumers who drive older cars. In a couple years, Mobileye offers "eyes off" on highways on consumer cars. It is only on highways but it is a safe and reliable "eyes off" system so the driver does not need to pay attention on long highway road trips. That would be very beneficial too for many consumers who take long trips. And then Waymo continues to scale and offers very smooth, safe, reliable robotaxi services in major cities and suburbs. That would also be very beneficial to people who can't drive or don't have access to a car in the city or just need a quick ride somewhere.

Wayve is still in the early stages. They have not removed the safety driver yet or deployed any commercial product, But I wish Wayve the best of luck. I hope they do well and I look forward to seeing just how far they can push E2E.

In this discussion several months ago, Anguelov says that E2E is not "there yet" but that things are moving in that direction because stacks are using fewer NNs and NNs are being consolidated. So it is entirely possible that one day, all AVs will be E2E, they just took different paths to get there.

 

[diplomat33]

I have watched hundreds of videos from Xpeng's CNGP and XNGP and Huawei's ADS (both door to door system).

XPENG rolls out NGP in Beijing to certain public testers:

"XPeng's (NYSE: XPEV) Advanced Driver Assistance System (ADAS) similar to Tesla's FSD (Full Self-Driving) is now available in Beijing, making the company the first to launch the feature in the Chinese capital city.

The XPeng feature, called City NGP (Navigation Guided Pilot), is open in Beijing for users participating in a public test and is currently available on Beijing's ring roads and major highways, according to a press release today.

In addition to public test users, the feature will soon be available for general users of the Max version of XPeng's flagship G9 and P7i with the Xmart OS 4.3.0 system update, it said.

XPeng P5's P-version models will also be able to use the City NGP feature in Beijing after upgrading to Xmart OS 3.5.0, according to the company."

XPeng rolls out Tesla FSD-like assisted driving feature in Beijing

City NGP feature is expected to cover dozens of additional cities within the year, XPeng said.

cnevpost.com

 

[scottf200]

8x speed clip of Wayve driving autonomously through busy London with vision-only end-to-end NN (ie one NN that takes vision in and outputs controls):

https://twitter.com/x/status/1669374996580298757

It does not mean that they have solved FSD. But it is a still an interesting demo of what end-to-end can do IMO. I can see why people see E2E as a promising approach.

I've been following Wayve for a while and they have impressive drives. Obviously some big backers and lot of powerful tech with Microsoft and NVidia scaleable cloud resources.

The ride with Gates was great.

 

[spacecoin]

I've been following Wayve for a while and they have impressive drives. Obviously some big backers and lot of powerful tech with Microsoft and NVidia scaleable cloud resources.

While I applaud the approach, I'm a Wayve sceptic. I think e2e solutions will prove to be too brittle and costly to change and deploy in the short term. It's a bit like a black box monolith in traditional software engineering. Explainability and debugging is critical to deploy in practice. Having a component based architecture helps a lot.

 

[ThomasD]

How many deliveries can Wayve due per hour? Are there multiple compartments in the vehicle to keep customer orders separate? If not how do they keep someone from taking an order that is not theirs?

 

[diplomat33]

How many deliveries can it due per hour?

We don't know. It would likely depend on how far it has to travel to each delivery.

Are there multiple compartments in the vehicle to keep customer orders separate? If not how do they keep someone from taking an order that is not theirs?

Yes, the customer orders would be kept separate. I am sure they have some type of system to make sure customers don't take the wrong order like giving them a special code that only unlocks the compartment with their order.

 

[stopcrazypp]

I voted Yes to both questions. If Tesla were to have done something similar and released something that was 1,000 times worse. You Tesla fans would be claiming it is the best thing since sliced bread and any delay would be killing people.

Oh wait they did and you guys responded exactly like that. They released one of the WORST and DANGEROUS software in FSD Beta when it clearly shouldn't be released and you guys praised and defended it. All you have to do is watch the old videos to see how horrific the software was at the times.

(You can make the case that version 10.69 or 11 should have been the version that is released)

Tesla never released a system where they had no driver in the car or told the Beta tester they didn't need to pay attention. In fact they had been strict on the attention criteria for FSD Beta release (kicking people off the program when they found it was violated), and had a previous safety score system (subject of a lot of angst in these forums). It was that strictness that is mostly credited to why FSD Beta incidents have been relatively minor (a lot of naysayers suggested accidents involving serious injuries or death would occur rapidly, using similar wording as you did). I personally expected statistics to eventually win and a fatal accident to eventually occur on FSD Beta, but so far it has passed the general vehicle fatality rate (1-1.3 fatalities per 100 million miles) without a fatality given 150 million miles have been travelled on FSD Beta already.
Tesla Full Self-Driving (FSD) Beta program surpasses 150 million miles
NHTSA estimates trafic fatalities increased 18.4% y-o-y in 1H 2021

That's the major difference between L2 and L4, so I'm surprised you missed that difference completely. No one here as argued FSD Beta should be allowed to operate without a driver in the seat in its current state.

In context, we are talking here about a L4 vehicle with no safety driver in the seat at all, and whether it is appropriate for it to not have one.

I have watched hundreds of videos from Xpeng's CNGP and XNGP and Huawei's ADS (both door to door system).
The state these systems were released in were A planet better than the state FSD Beta was released in.
The same thing was the case with the release of NOA. It was horrific and you guys praised and defended it.
Other systems equivalent at release were no where near that state.

Lol, this argument reminds me so much about the media and naysayers back when it was released talking about how untested AP was when it came out, when the state of the art "fully tested" systems from other automakers like Mercedes would happily run the car into oncoming traffic, but you hardly heard a peep about that.
17.26.76

It demonstrates that Tesla is willing to release software that are not Ready.
What makes you think they won't continue this trend when they finally do robo-taxi?

As above, they had been fairly risk adverse in their FSD Beta release, so why would robo-taxi be different? They are the company with NHTSA breathing down their neck the most.

 

[Bladerskb]

Tesla never released a system where they had no driver in the car or told the Beta tester they didn't need to pay attention. In fact they had been strict on the attention criteria for FSD Beta release (kicking people off the program when they found it was violated), and had a previous safety score system (subject of a lot of angst in these forums). It was that strictness that is mostly credited to why FSD Beta incidents have been relatively minor (a lot of naysayers suggested accidents involving serious injuries or death would occur rapidly, using similar wording as you did). I personally expected statistics to eventually win and a fatal accident to eventually occur on FSD Beta, but so far it has passed the general vehicle fatality rate (1-1.3 fatalities per 100 million miles) without a fatality given 150 million miles have been travelled on FSD Beta already.
Tesla Full Self-Driving (FSD) Beta program surpasses 150 million miles
NHTSA estimates trafic fatalities increased 18.4% y-o-y in 1H 2021

That's the major difference between L2 and L4, so I'm surprised you missed that difference completely. No one here as argued FSD Beta should be allowed to operate without a driver in the seat in its current state.

In context, we are talking here about a L4 vehicle with no safety driver in the seat at all, and whether it is appropriate for it to not have one.

Lol, this argument reminds me so much about the media and naysayers back when it was released talking about how untested AP was when it came out, when the state of the art "fully tested" systems from other automakers like Mercedes would happily run the car into oncoming traffic, but you hardly heard a peep about that.
17.26.76

As above, they had been fairly risk adverse in their FSD Beta release, so why would robo-taxi be different? They are the company with NHTSA breathing down their neck the most.

This is irrelevant.

I'm talking about door to door L2 system on release versus door to door L2 system on release.
We have thousands of videos of FSD Beta near release and we have thousands of videos of Huawei ADS and others near release.
You completely ignored the comparison. Every system that Tesla has released has been absolutely trash on release.
This is undeniable, but not to Tesla fans.

 

[diplomat33]

I've been following Wayve for a while and they have impressive drives. Obviously some big backers and lot of powerful tech with Microsoft and NVidia scaleable cloud resources.

The ride with Gates was great.

You should find this interesting since you follow Wayve. They just tweeted this very interesting thread on their new generative AI model called GAIA-1. It is a self-supervised AI that can generate the next frame of a video based on the current frame. Basically they feed it a ton of video data from the real world and it can learn what vehicles, pedestrians etc do in different situations. It can be used to train their end-to-end to predict what other vehicles or pedestrians might do and plan the right action. The thread has lots of cool examples.

https://twitter.com/x/status/1669671852938088449

https://twitter.com/x/status/1669671861288939520

 

[diplomat33]

While I applaud the approach, I'm a Wayve sceptic. I think e2e solutions will prove to be too brittle and costly to change and deploy in the short term. It's a bit like a black box monolith in traditional software engineering. Explainability and debugging is critical to deploy in practice. Having a component based architecture helps a lot.

That is the same argument Anguelov gave against e2e. I am skeptical too. It just seems like a monumental task to train an e2e AI to intelligently and reliably drive everywhere. But I am trying to be open minded about it. They seem to be doing really good work and AI is making great strides. Heck, maybe they don't achieve unsupervised L5. But even a vision-only e2e L2 door-to-door system would still be very valuable IMO. It is certainly worth trying.

EDIT: I wonder how wayve approach will scale to different countries with different road rules. For example, the US and UK don't drive on the same side of the road. That's a big difference. With e2e, I feel like they would need to retrain their entire system from scratch to work in the US. If so, that would be a disadvantage to e2e.

 

[stopcrazypp]

Looks like BMW beat Mercedes to approval of higher L3 speeds in Germany (130km/h is approximately 80mph):

BMW 5 Series is the First Car with German Approval for 130 KM/H Highway Assist

The new 2024 BMW 5 Series is the first car to get Germanys approval for its semi-autonomous motorway assist.

www.bmwblog.com

 

[ZeApelido]

OK let's summarise why HW3/4 isn't good enough.

1. Guessing range from a 2D image with only semantic cues is not safe enough, especially when there are few reference objects and at night. Tesla's been running into things (first responders, motor cyclists) at night. It is safer to physically measure the distance, so why wouldn't you?

2. If you travel at 60 mph, and there is smoke or fog on the road, cameras cannot see. If the cameras are blinded by oncoming traffic or the sun, they cannot see. There are too many failure modes for cameras for them to be safe enough to trust with your life.

3. Even if Tesla would add more cameras (in the a-pillars and the front for example) why wouldn't you want to use more sensor modalities to make a 10x safer product for almost zero cost? Tesla won't even bother to add the HW4 radar to 3/Y. Tesla picks lowering cost over improving FSD safety every time.

1. It's not "guessing", it's utilizing cues from structure in motion or learned priors, much like humans do. Can you look through a camera and determine object distance and segmentation? If so, a computer can theoretically do it, assuming unlimited compute and ML advances. Tesla's perception stack is not a party-trick. That doesn't make it as accurate in depth estimation as Lidar, but there's nothing fundamentally flawed about it that can't be overcome with time to match near human performance.

2. So who said you have to do better than humans in these scenarios, especially when starting out?

3. Sure, when the rest of their FSD stack is good enough to actually deploy as Robotaxis and not L2. Then it makes sense to add other sensors. Why waste the money now?

Cruise's and Waymo's solutions generalize better than you think. They can add a new city extremely fast so that's not a scaling bottle neck anymore - other things like permits, building up operations and service locations is. Tesla isn't even playing the same game. Robotaxi is a completely different use case than driver assistance systems.

Extremely tenuous statements. No proof their systems fully generalize over different cities with different behaviors, that's not how data science / ML works. And their slow roll outs only further assert that it is indeed a problem. Show me where they have deployed to a new city "fast".

If the software was robust, once the mapping is finished, deployment in a new city should be over in a few weeks. Full SF coverage across both night and day didn't even happen anywhere close to that.

 

[diplomat33]

Extremely tenuous statements. No proof their systems fully generalize over different cities with different behaviors, that's not how data science / ML works. And their slow roll outs only further assert that it is indeed a problem. Show me where they have deployed to a new city "fast".

Well, according to Dolgov, the Waymo Driver does generalize well across many cities. The proof is that they were able to scale simultaneously in both SF and Phoenix and when they "dropped" the Waymo Driver in LA, it showed great performance from the get-go.

https://twitter.com/x/status/1606055081358413824

If the software was robust, once the mapping is finished, deployment in a new city should be over in a few weeks. Full SF coverage across both night and day didn't even happen anywhere close to that.

The software is very robust:

https://twitter.com/x/status/1606054223111544832

Deployment is fast. Waymo can start autonomous driving in a new city very quickly. Just map and go. They did it in LA. What takes time is going from deployment to a public driverless service. That's because Waymo needs to verify safety first before they will let the public ride in a Waymo with no safety driver. And Waymo cannot assume it is safe in a new city, just because it was safe in a previous city since each city is different. So, Waymo needs to drive around in each new city to prove safety in the new city. That takes time. So it's the safety validation that slows Waymo down from launching a public driverless service in a new city. Now, as the Waymo Driver generalizes better, the validation time will get shorter. Waymo is also trying to shorten the validation time with better simulations but you will always need real world driving to prove safety. You can't get around that.

 

[spacecoin]

1. It's not "guessing", it's utilizing cues from structure in motion or learned priors, much like humans do. Can you look through a camera and determine object distance and segmentation? If so, a computer can theoretically do it, assuming unlimited compute and ML advances. Tesla's perception stack is not a party-trick. That doesn't make it as accurate in depth estimation as Lidar, but there's nothing fundamentally flawed about it that can't be overcome with time to match near human performance.

No one doubts that "a computer can theoretically do it, assuming unlimited compute and ML advances".
But we're talking about a product that is on cars today, not some fictional product in 40 years. In 40 years the race is already over.

Relying on semantic cues alone is not safe enough to remove the driver under challenging conditions such as low visibility. That is just a fact at the present time.

2. So who said you have to do better than humans in these scenarios, especially when starting out?

Literally everyone that works in AVs says that. For robotaxi to gain societal acceptance they need to be a lot more reliable than the average human driver. They also need to drive better.

3. Sure, when the rest of their FSD stack is good enough to actually deploy as Robotaxis and not L2. Then it makes sense to add other sensors. Why waste the money now?

So you're saying FSD is a scam or a bait and switch? Elon said Level 5 autonomy on existing cars multiple times. He also said there will be no HW3 upgrades to anything better.

Extremely tenuous statements. No proof their systems fully generalize over different cities with different behaviors, that's not how data science / ML works. And their slow roll outs only further assert that it is indeed a problem. Show me where they have deployed to a new city "fast".

Let's follow up in 12 months. Scaling up a taxi operation takes time. It's not just about technology. It's about working with the community to strengthen your brand, build acceptance and getting permits, getting ops centers and hubs set up etc. Hire people to manage and clean the fleet and so on. These things take time.

If the software was robust, once the mapping is finished, deployment in a new city should be over in a few weeks. Full SF coverage across both night and day didn't even happen anywhere close to that.

No. How do you set up a taxi operation (assuming you already have drivers) in a "few weeks" in a new city?

 

[ZeApelido]

What takes time is going from deployment to a public driverless service. That's because Waymo needs to verify safety first before they will let the public ride in a Waymo with no safety driver. And Waymo cannot assume it is safe in a new city, just because it was safe in a previous city since each city is different. So, Waymo needs to drive around in each new city to prove safety in the new city.

Safety validation is all we are talking about in the end. It's the performance in the real-world. In Waymo's case, when they begin safety validation in a new city they are doing so presumably without much training data in that city.

Is your assertion that Waymo does not make any changes to their software over the course of this safety validation phase? That they do not use the data collected to modify their software?

No one doubts that "a computer can theoretically do it, assuming unlimited compute and ML advances".
But we're talking about a product that is on cars today, not some fictional product in 40 years. In 40 years the race is already over.

Relying on semantic cues alone is not safe enough to remove the driver under challenging conditions such as low visibility. That is just a fact at the present time.

Your descriptive "semantic cues" isn't a full description of what is happening and is an attempt to diminsh the technical achievement. There is a real statistically learned physics-based algorithm behind this...much like the human visual cortex.

Your 40 years trajectory is laughable considering how things have trended last 5-10 years. I'm not arguing the current perception solution is definitely good enough to allow better than human driving performance, but I don't see any big roadblocks stopping it from being that within a few years. Heck even 4x the resolution andand compute power in HW4 and resultant 4x in depth map resolution will help.

If you've paid attention to the FSD videos, perception is almost never the issue. This was actually surprising even to me.

So you're saying FSD is a scam or a bait and switch? Elon said Level 5 autonomy on existing cars multiple times. He also said there will be no HW3 upgrades to anything better.

Oh yeah, I could care less about HW3. Better camera resolution & placement and higher compute power will make HW4.0 superior for FSD within a year or 2. HD Radar can improvement reliability. Musk needing to stop FSD claims doesn't take away from the technology development.

Let's follow up in 12 months. Scaling up a taxi operation takes time. It's not just about technology. It's about working with the community to strengthen your brand, build acceptance and getting permits, getting ops centers and hubs set up etc. Hire people to manage and clean the fleet and so on. These things take time.

No. How do you set up a taxi operation (assuming you already have drivers) in a "few weeks" in a new city?

These are all true, and can easily mask that the software still needs to be tweaked in that timeframe to work well in those cities.

I'm sure that when Waymo starts driverless rides in LA, it will span a wide region of LA, right? Not a small section, right?