Autonomous Car Progress

[replicant]

I was cycling yesterday in downtown Paris and followed a Model Y w/ NL plates crossing one of the busiest place during rush-hour (Place de Clichy at ~6PM).

I noticed it had a metal structure on the roof that held what looked like small sensors. At least two were pointing to the rear, and were close to the lengthwise axis of the car. I assume two others where pointing forward and maybe two others on each side.

I guess the car is operated by Tesla (given the NL plate and the model) and that the sensors were Lidar. They looked very much like these

https://twitter.com/x/status/1395356317410287617

Why would they calibrate the Y in such busy places, at the worst possible time? Could this be for FSD localization?

 

[drtimhill]

That's one reason why companies geofence. It helps manage the workload. But remote operators and geofence are only needed now because autonomous driving is not quite at 99.999999% yet. As autonomous driving gets better, we will need remote operators less and less. Eventually, autonomous driving will be so good that we can trust them to drive without any remote operators and no geofence.

Not sure I agree with this logic. My understanding is the primary driver of geofencing is detailed HD maps etc necessary for these cars to operate at all, though no doubt the restricted operational area is a help to the remote operators.

 

[diplomat33]

Not sure I agree with this logic. My understanding is the primary driver of geofencing is detailed HD maps etc necessary for these cars to operate at all, though no doubt the restricted operational area is a help to the remote operators.

Contrary to popular opinion, building detailed HD maps is not that difficult. AV companies can build detailed HD maps of large areas relatively easily. They just need to drive around a lot to collect data. Supercomputers then crunch the data and automatically generate the maps. HD maps might part of the reason but I don't think it would be the main reason for geofencing. I think there are other reasons that would be bigger factors. Testing is probably the biggest reason. A smaller area will make testing easier since there will be less edge cases, quicker time to get to an AV if it gets into trouble, you don't need as many vehicles which also reduces maintenance, reduces the need for as many remote operators, reduces risk of an accident. I feel those reasons would be a bigger reason to geofence than HD maps.

 

[EVNow]

Contrary to popular opinion, building detailed HD maps is not that difficult. AV companies can build detailed HD maps of large areas relatively easily.

Not according to Cruise. I've linked before.

 

[diplomat33]

Not according to Cruise. I've linked before.

I think you are misunderstanding what Cruise said. I don't think Cruise said that building HD maps is super hard. But show me the quote again if you think I am wrong.

 

[EVNow]

I saw that quote. I think you are misunderstanding what Cruise said. I don't think Cruise said that building HD maps is super hard. But show me the quote again if you think I am wrong.

It is not a question of HD Maps being "super hard". They are simply not an option if you want millions of miles to be covered and updated constantly. See what GM says as well.

Anyway, Cruise basically said they compromised on several items included small ODD and HD Maps to get to high level of error free operation quickly.

 

[Dewg]

There's an easy solution to HD maps - every car is loaded with cameras. So every car driving every street should be updating their maps pretty consistently. Your car sees construction cones as it's driving? Bam, the maps are updated for the next car that comes by in an hour or so, so that car knows the cones are there. Same concept as Waze - crowd source the map data!

 

[drtimhill]

Contrary to popular opinion, building detailed HD maps is not that difficult. AV companies can build detailed HD maps of large areas relatively easily. They just need to drive around a lot to collect data. Supercomputers then crunch the data and automatically generate the maps. HD maps might part of the reason but I don't think it would be the main reason for geofencing. I think there are other reasons that would be bigger factors. Testing is probably the biggest reason. A smaller area will make testing easier since there will be less edge cases, quicker time to get to an AV if it gets into trouble, you don't need as many vehicles which also reduces maintenance, reduces the need for as many remote operators, reduces risk of an accident. I feel those reasons would be a bigger reason to geofence than HD maps.

Well yes, but "not that difficult" followed by your explanation sounds, while not "difficult", like an expensive and time-consuming operation. And dont forget that these maps are out of date the moment they are published, so in fact this is an ongoing process.

it's obvious to me the "HD maps" (Waymo) vs "smart cars" (Tesla) approaches are going to converge over time. What is every Tesla with FSD beta really doing? Building its own local HD map (or as best as it can get) on-the-fly as it drives. Ok, so why shouldn't the car share the information about the layout of an intersection back to Tesla? If every car does that from all the different angles and approaches to an intersection, Tesla are suddenly in the process of the fleet doing that mapping purely as a side-effect of driving. Crowd-sourced on-the-fly HD mapping as it were, with much of the "super-computer" being the combined compute power of all the cars. Add some clean-up at the mother ship, and you have cars that both map their environment and can benefit from all the other cars that have already encountered that intersection.

 

[diplomat33]

It is not a question of HD Maps being "super hard". They are simply not an option if you want millions of miles to be covered and updated constantly. See what GM says as well.

Ok but ride-hailing does not need to cover millions of miles. So it is not relevant to robotaxi companies like Waymo and Cruise that are focused on deploying robotaxi services. It is only relevant if you are trying to deploy consumer cars that can drive autonomously everywhere. Of course, Mobileye has AV maps which are less detailed than the HD maps that Waymo or Cruise use and those AV maps do work over millions of miles.

So HD maps is not why they geofence. They geofence because their goal is deploying robotaxis that only need to work in a limited service area and testing and deploying those robotaxis is easier in a smaller area.

I just think you are twisting their words. They are saying they chose to geofence and use HD maps in order to make "solving FSD" easier. That's not the same as saying they have to geofence because they use HD maps.

Anyway, Cruise basically said they compromised on several items included small ODD and HD Maps to get to high level of error free operation quickly.

Because the goal for AV companies like Waymo and Cruise is to deploy a viable and profitable robotaxi service as quickly as possible. To do that, they need to achieve a high level of error free operation as quickly as possible. So for companies like Cruise and Waymo, achieving that high level of error free operation is more important than the size of the ODD. But again, that does not mean that HD maps is the reason why they geofence. Achieving a high level of error free operation is the reason why they geofence. HD maps help them achieve that goal.

There's an easy solution to HD maps - every car is loaded with cameras. So every car driving every street should be updating their maps pretty consistently. Your car sees construction cones as it's driving? Bam, the maps are updated for the next car that comes by in an hour or so, so that car knows the cones are there. Same concept as Waze - crowd source the map data!

Yes. And that is what Mobileye is doing now. They call the maps, "AV maps". They are crowdsourced from cameras only from the existing fleet of Mobileye powered vehicles on the road equipped with a front facing camera.

 

[drtimhill]

Ok but ride-hailing does not need to cover millions of miles. So it is not relevant to robotaxi companies like Waymo and Cruise that are focused on deploying robotaxi services. It is only relevant if you are trying to deploy consumer cars that can drive autonomously everywhere. Of course, Mobileye has AV maps which are less detailed than the HD maps that Waymo or Cruise use and those AV maps do work over millions of miles.

So HD maps is not why they geofence. They geofence because their goal is deploying robotaxis that only need to work in a limited service area and testing and deploying those robotaxis is easier in a smaller area.

hmmm I think this is a bit post-event revisionist. Waymo most certainly were on a mission to develop a consumer self-driving car (or at least the technology stack they could sell to other car makers), and only pivoted to geofenced ride hailing when that original goal proved harder than they anticipated. So saying they dont need millions of miles in HD maps because they are only doing ride hailing is kinda backward logic.

 

[drtimhill]

Sorry, I think you are getting history wrong here. If you are referring to Waymo developing their consumer car ADAS that ironically was going to be called "autopilot", Waymo explained that the reason they dropped it, is because the test subjects were not paying attention. Waymo felt like a self-driving system that relied on driver attention was too risky. So they pivoted to L4 that does not require driver attention. They did not pivot to L4 because consumer self-driving cars was too difficult.

But that's my point .. I didn't say it had to be a technical reason for the pivot (though I'm suspicious this stated reason, which, yes, I do remember, was not the only reason). And my point remains .. the only way they could (quickly) figure out an L4 system was with HD maps, and the limitation here was the lack of such maps, so they were more or less forced to choose a geofenced ride-hailing approach. Sure, they put a spin on it (gasp!), but the reality was they were more or less forced into the pivot.

 

[SVMike]

There's an easy solution to HD maps - every car is loaded with cameras. So every car driving every street should be updating their maps pretty consistently. Your car sees construction cones as it's driving? Bam, the maps are updated for the next car that comes by in an hour or so, so that car knows the cones are there. Same concept as Waze - crowd source the map data!

I think that whatever data (maps or otherwise) that are sent to a car need to be curated by the company if it is going to be used as anything other than a "displayed hint" or "suggestion" by the car.
Otherwise we are just creating an opportunity for a few people to setup a stupid/dangerous situation, then they drive some cars by it, then all subsequent cars do something stupid (like drive to the edge of a ditch) and cause a traffic jam, etc. Once someone did this and the hilarious video went viral you'd have copycats everywhere.

 

[diplomat33]

And my point remains .. the only way they could (quickly) figure out an L4 system was with HD maps, and the limitation here was the lack of such maps, so they were more or less forced to choose a geofenced ride-hailing approach. Sure, they put a spin on it (gasp!), but the reality was they were more or less forced into the pivot.

I think you are putting a spin on things.

Waymo figured out L4 with a combination of HD maps, camera vision, radar and lidar. So I would not say that HD maps is the only way they could figure out L4. Again, HD maps help make L4 better. But Waymo does not rely on HD maps alone to do L4. Waymo uses camera vision, radar and lidar. And the fact is that HD maps account for a very small part of the autonomous driving stack. The vast majority of the driving decisions come from the perception (camera, radar and lidar), prediction and planning stack.

And Waymo could have gone the Tesla route and deployed a L2 "FSD" system that requires driver supervision. In fact, Waymo had a very capable L2 system. Waymo chose to drop the L2 system because they felt it was too risky since humans are not good at supervising. So they decided to only do autonomous driving that does not require driver supervision. But autonomous driving that does not require any driver supervision that works everywhere is unrealistic. So L4 was their only realistic option if they were only going to do autonomous driving that does not require driver supervision. And at the point when they decided to go the L4 route, yes, geofenced ride-hailing was a logical direction. But they were not forced into it.

There is a big difference between "they couldn't solve self-driving on consumer cars because it was too difficult so they had to do geofence with HD maps" and "they chose not to do self-driving with driver supervision because of lack of driver attention so they turned to geofenced self-driving that could be done with no driver supervision."

 

[Daniel in SD]

I think that whatever data (maps or otherwise) that are sent to a car need to be curated by the company if it is going to be used as anything other than a "displayed hint" or "suggestion" by the car.
Otherwise we are just creating an opportunity for a few people to setup a stupid/dangerous situation, then they drive some cars by it, then all subsequent cars do something stupid (like drive to the edge of a ditch) and cause a traffic jam, etc. Once someone did this and the hilarious video went viral you'd have copycats everywhere.

So someone would reconfigure the road to get the HD map updated then switch it back? I bet most the time the car would be able to detect that reality no longer matched the map. Recognizing ditches is not hard (for LIDAR at least).

 

[Doggydogworld]

Dirty Tesla has a new clip with the latest Beta. There are a lot of tricky areas it handled well. A few things I noticed.
2:45 - gets in right lane but needs to turn left. Recovers poorly.
4:35 - deftly slows for a car that illegally turns in front of him without warning
5:30 - handles a series of tricky turns well
14:00 - runs a red light (Elon will pay the ticket, right?)
14:20 - gets in lane that has construction coming up, recovers late but well
15:35 - gets in wrong left turn lane, recovers pretty well
17:00 - wanders a bit in parking lot, but not dangerous

 

[SVMike]

So someone would reconfigure the road to get the HD map updated then switch it back? I bet most the time the car would be able to detect that reality no longer matched the map. Recognizing ditches is not hard (for LIDAR at least).

Tesla shorts have gone to greater lengths over the past few years than this.

 

[Daniel in SD]

Tesla shorts have gone to greater lengths over the past few years than this.

Manslaughter?

 

[ThomasD]

Aiming to make car ownership a thing of the past

Inside Zoox: The robot vehicle totally changing transportation

Take an exclusive ride in the vehicle with no steering wheel that’s aiming to make car ownership a thing of the past.

www.freethink.com

 

[Daniel in SD]

Aiming to make car ownership a thing of the past

Inside Zoox: The robot vehicle totally changing transportation

Take an exclusive ride in the vehicle with no steering wheel that’s aiming to make car ownership a thing of the past.

www.freethink.com

That article has so many errors regarding SAE levels...
Also, the reason consumers can't ride in Zoox vehicles has nothing to do with regulations.

 

[diplomat33]

Looks like Cruise got their permit to start charging money for their driverless rides in SF:

https://twitter.com/x/status/1532506250440605696