Tesla Autonomy Day April 22nd

[neroden]

My biggest takeaway from the presentation is that my intuition strongly told me Tesla was doing some particular things wrong with the computer assisted driving in the past (various particular approaches they were using), and at least one or two of those mistakes were actually identified in Monday's presentation as "we tried that and it was wrong and now this is what we're doing" during this presentation, and what they described as their new improved version is much closer to what my intuition originally said should be done years ago when I had that intuition. That tells me that, among other things, every one of their (important-enough) mistakes they will eventually fix; they don't just stay stuck in one way of doing things.

Well, I agree entirely!

Since I already know some of the things they're doing wrong (just as you did!) I can guesstimate how many years the trial-and-error over those things are going to add to the timeline.

I saw a lot of improvement and progress in the self driving work, and a lot of this talk was to show people who know better that they are indeed on that path, not flailing about.

OK, I guess that's a possible reason for having the day? It was a nice set of lectures, and it did give me a good impression of where they are on the path.

I'm not super familiar with most of the other companies' efforts because they don't give a lot of talks. Cruise definitely has the right attitude of collecting edge cases, and I hope they can abandon LIDAR; Waymo seems to have locked itself into a totally wrong-headed approach; and Uber, well, no chance, they just don't care about human life.

 

[neroden]

Even the AP1 hardware was an example of this: they claimed it was enough for FSD, but when it turned out not to be, they totally abanoned that presumption (but as you noticed, the lawyers needed to pull the rug out from underneath those owners inperceptably slowly, so that was painful for those of us with long memories to watch, but they seem to have gotten away with it, including because they believed Elon when he said they were paying for the development but Tesla will still need to figure it out (which as I point out here seems true) so they gave him that slack).

I question how much Tesla will pay in refunds for that. Anyone with AP1 hardware who ordered "FSD" deserves a full refund; I hope Tesla's just paying the refunds out. From previous information, it seemed like it wasn't a material amount of money because not that many people bought the vaporware.

 

[MP3Mike]

I question how much Tesla will pay in refunds for that. Anyone with AP1 hardware who ordered "FSD" deserves a full refund; I hope Tesla's just paying the refunds out. From previous information, it seemed like it wasn't a material amount of money because not that many people bought the vaporware.

Nobody bought it as FSD wasn't an option on AP1, so there is nobody that would qualify for that refund.

 

[Cosmacelf]

Nobody bought it as FSD wasn't an option on AP1, so there is nobody that would qualify for that refund.

That’s right. AP1 never had the ability to purchase FSD. I have one of the later versions of AP1 on my car, and there was never any mention of FSD. That only became an option with the V2 AP hardware that included 8 cameras. AP1 only uses a single forward facing camera.

 

[Ben D]

Elon killed two things today, all hopes for AP1 with his “buying a car without FSD today is like buying a horse” statement and somehow killed amphibians with a statement so profound I am truly unable to understand what he meant. R.I.P amphibians...

I took this as Elon comparing FSD to evolution of humans from amphibians. The amphibians are the ones that crawled out of the water. and became the first land animals. Humans are the latest in the line of land animals - humans who are now making cars drive by themselves so they don't even have to walk on land.... Beat that amphibians...

 

[Ben D]

So, discussion of the fantasy business model for the fantasy robotaxis. Which Musk delusionally thinks they'll have next year. There is no point in even listening to this delusional nonsense.

They won't. For those who doubt me, I'd like to point out that my predictions have been substantially better than Musk's on this stuff for the past several years, and you can go back and check my record.

If you suppose that the fictional robotaxis will exist (which isn't happening any time soon), he makes further mistakes:

The idea that you can get the average car up to 60 hours a week by using it as a robotaxi is incorrect as it does not sufficiently account for rush hour behavior. It depends on what city you're in, actually (how much shift-work there is versus spread-out traffic).

Glad they're designing the car to last forever, though.

Musk continues with his delusional fantasy of no steering wheels and no pedals. He proposed 2 years from now. He's flat out wrong, mark my words.

I know people don't want to hear this. Self-driving is a very popular fantasy, just as it was two years ago (when I threw cold water on this), and just as it was in the 1950s. But if you're investing -- get a clue and pay attention to reality.

I know little about computers but a fair bit about physics and do a lot of driving. I also deal with politicians and regulators a lot. It sounded to me like the hardware was adequate and the sensor suite was enough to do the job. I think Elon has compiled a team of very smart people at the pointy end of the right science fields, and with a fleet of 500,000 cars to learn from to fine tune their software they will have access to more real world information than anyone on earth. If the neural network can be trained to learn from this database, given it is very likely their data is increasing exponentially, they will have access to temporal trend data which can be interpolated to indicate when the software will be good enough to be X times safer than the average human. I think they have done this and that is how they arrived at their timeline.

The key here which I think some people are overlooking, is that it doesn't have to be perfect, just much better than the average human, who in reality is quite a poor inattentive driver with slow reaction times, only 2 eyes and ears, little understanding of vehicle dynamics and other stuff. From what I saw and heard, Tesla have what it takes. They have more real world data than anyone else because of their ever expanding fleet which will help fill in the tail end, that data will be used to inform regulators (who won't listen to simulations), and that will be enough to allow FSD to happen somewhere (it may be in Europe or China first) within the timeframes they are speaking of. Sure, there will still be auto accidents with FSD, but once its obvious that the accident rate for the computer driven car is significantly less than that for real people, there is no need for regulators to wait, and indeed there will be pressure for them to implement.... So even if the self driving car is only as good as an excellent driver (e.g. Neroden?) in driving ability, that would probably be enough to reduce accident rates by over 75% by the sounds of it, which most regulators would jump at in a heartbeat..

 

[Buggle]

I know little about computers but a fair bit about physics and do a lot of driving. I also deal with politicians and regulators a lot. It sounded to me like the hardware was adequate and the sensor suite was enough to do the job. I think Elon has compiled a team of very smart people at the pointy end of the right science fields, and with a fleet of 500,000 cars to learn from to fine tune their software they will have access to more real world information than anyone on earth. If the neural network can be trained to learn from this database, given it is very likely their data is increasing exponentially, they will have access to temporal trend data which can be interpolated to indicate when the software will be good enough to be X times safer than the average human. I think they have done this and that is how they arrived at their timeline.

The key here which I think some people are overlooking, is that it doesn't have to be perfect, just much better than the average human, who in reality is quite a poor inattentive driver with slow reaction times, only 2 eyes and ears, little understanding of vehicle dynamics and other stuff. From what I saw and heard, Tesla have what it takes. They have more real world data than anyone else because of their ever expanding fleet which will help fill in the tail end, that data will be used to inform regulators (who won't listen to simulations), and that will be enough to allow FSD to happen somewhere (it may be in Europe or China first) within the timeframes they are speaking of. Sure, there will still be auto accidents with FSD, but once its obvious that the accident rate for the computer driven car is significantly less than that for real people, there is no need for regulators to wait, and indeed there will be pressure for them to implement.... So even if the self driving car is only as good as an excellent driver (e.g. Neroden?) in driving ability, that would probably be enough to reduce accident rates by over 75% by the sounds of it, which most regulators would jump at in a heartbeat..

Europe will be last, not first. First will be California if I had to guess. If not there then maybe Nevada another southwest state. Somewhere in the US where Tesla has a good reputation & where there gov't regulators are willing to try new things.

 

[neroden]

The key here which I think some people are overlooking, is that it doesn't have to be perfect, just much better than the average human,

Based on the psychology which I know from other areas, I think it has to be 50 to 100 times better than the average human. People have a "double standard". I have three sources for data on this: one is the standards to which automated trains and automated airplanes have been held.

The other: there are different economic analyses of what people are willing to pay to prevent an avoidable death, from *different fields* with *different types of avoidable deaths* -- and they're willing to pay a lot less to avoid a death when there's a human to blame, then they are willing to pay to avoid "death by machine". You can look this up yourself.

The final one: people will voluntarily choose more dangerous options if they feel that they're in "control". So people who are rattled by airplane crashes and train crashes will choose to drive, even though it's objectively several times more dangerous. This is also easy enough to look up.

So even if the self driving car is only as good as an excellent driver (e.g. Neroden?) in driving ability, that would probably be enough to reduce accident rates by over 75% by the sounds of it, which most regulators would jump at in a heartbeat..

Trouble is I'm quite sure that isn't actually enough. Regulators, like everyone else, demand a higher standard from robots than from humans. Being somewhat more data based than the average human, but still with an eye towards public opinion, they may only demand 10 or 20 times better. Or, if they're particularly paranoid about public opinion, they might even demand a higher standard than the public demands.

This is going to take a long time.

Now, humans PLUS robots ("driver assistance") is held to a *much* lower standard, so that can get approved quickly. A vehicle where the machine does nearly all the driving but there is someone to blame in the driver's seat will be held to a far, far lower standard. That's why this is the path forward for approval.

Railroad train drivers have basically zero ability to prevent a crash if they see an obstacle ahead. But people still are uncomfortable having the trains have no drivers, even if all the driver does is sit around with the ability to hit the big red brake button. People are very unhappy with dying in an airplane crash, but happier if they can blame the pilot than if the *airplane* failed. It goes on and on.

In the next several years, I won't believe a business model which depends on having nobody in the driver's seat.
A business model where the person in the driver's seat doesn't have to do anything except watch out for weird stuff and hit the "stop" button... that I can believe.

----
I wonder. How much would you have to pay a "Tesla Network Driver" for the job of taking the blame if anything goes wrong? Maybe less than Uber drivers are paid now? The job would become similar to sitting around and looking out the window of a bus, sightseeing. (No reading, though! Bad Uber!)

 

[EVNow]

I wonder. How much would you have to pay a "Tesla Network Driver" for the job of taking the blame if anything goes wrong? Maybe less than Uber drivers are paid now? The job would become similar to sitting around and looking out the window of a bus, sightseeing. (No reading, though! Bad Uber!)

Tesla will not hire drivers.

Tesla has to just convince regulators to allow the car to drive to the customer location and to the parking lot once the customer is done. During the trip the customer is in charge i.e. they decide how much of automation to use.

They can do this indefinitely until the regulators allow full FSD and even then, customer can still decide whether they drive (with EAP) or let FSD take over. All this until they remove the steering wheel.

ps :

BTW, I expect many if not most customers to slowly get used to various driver assistance features - and then steering and then NOA and be ready to completely give up control by the time regulators allow it.

It took me quite a few days to try even TACC, but once I started that, I quickly moved to lane keeping and lane changing. Next step for me is NOA.

 

[powertoold]

So there is a hole in Tesla's approach to FSD. They claim to use radar to tag the distances of cars in front of them; however, they have no such system to tag distances of cars on the right and left. So when you're making one of those unprotected left turns, how is the one camera on the b pillars going to figure out the distances of all the cars cross traffic?!

 

[MP3Mike]

So there is a hole in Tesla's approach to FSD. They claim to use radar to tag the distances of cars in front of them; however, they have no such system to tag distances of cars on the right and left. So when you're making one of those unprotected left turns, how is the one camera on the b pillars going to figure out the distances of all the cars cross traffic?!

Actually they said that they used vision to calculate distances of cars in front and used the radar to verify that it was accurate and as a backup. So vision only is enough to get distance.

 

[Cosmacelf]

So there is a hole in Tesla's approach to FSD. They claim to use radar to tag the distances of cars in front of them; however, they have no such system to tag distances of cars on the right and left. So when you're making one of those unprotected left turns, how is the one camera on the b pillars going to figure out the distances of all the cars cross traffic?!

You can estimate approx. distance using a camera. Humans do it all the time with our eyes, and no, we don’t use parallax much. Someone, I forget who, showed a 3D reconstruction of a scene using only a single camera. Basically, you are using past history and knowledge of how the world looks to determine distance. I don’t know if they are using frame by frame comparison to determine motion, but that is also something they could do.

 

[Dan Detweiler]

Actually they said that they used vision to calculate distances of cars in front and used the radar to verify that it was accurate and as a backup. So vision only is enough to get distance.

I believe they explained that they are triangulating the various camera view to determine distances around the car. Did I totally miss that?

Dan

 

[powertoold]

Someone, I forget who, showed a 3D reconstruction of a scene using only a single camera. Basically, you are using past history and knowledge of how the world looks to determine distance. I don’t know if they are using frame by frame comparison to determine motion, but that is also something they could do.

It was Karpathy and he used that point to say that the 3 front cameras are like two human eyes, being able to use figure out distance by using parallax. Then Karpathy even pointed to the example of using radar to help with tagging distances in front. They totally left a hole in their approach by not talking about determining distances left and right of the car...

I'm sure they thought about it, but for the sake of the presentation, it was left out.

 

[Ben D]

I wonder. How much would you have to pay a "Tesla Network Driver" for the job of taking the blame if anything goes wrong? Maybe less than Uber drivers are paid now? The job would become similar to sitting around and looking out the window of a bus, sightseeing. (No reading, though! Bad Uber!)

That is what Tesla owners will be doing for Tesla for free once FSD software comes online. I can see that is the initial step, private owners becoming "ride share sitters" collecting their 70% of the fare (Tesla getting 30% was it ?) and collecting data on FSD mode until regulators get enough data to show that tesla robotaxis can do it safely themselves, then the owners are free to let their car off the leash by itself and Tesla then puts its own cars on the road. It will be a straight simple statistical comparison of accident risk , and if the FSD cars have lower accident rates than average Joe, and they are cheaper to boot, there will be a lot of pressure to allow the new technology from the usual left wing sources.... If tesla gets FSD first if you are a ride share driver or owned a ride share fleet or a taxi fleet, buying teslas would become a pretty compelling business case.

 

[Ben D]

Europe will be last, not first. First will be California if I had to guess. If not there then maybe Nevada another southwest state. Somewhere in the US where Tesla has a good reputation & where there gov't regulators are willing to try new things.

If I had to read the tealeaves, I think it will occur first somewhere outside the USA, due to the influence of Wall Street on your regulators and interference from entrenched interests. Somewhere in Asia first I bet, then Europe will be close behind but just pipped to the post by somewhere in the USA as some west coast state sees sense supporting the locally made product....

 

[Buggle]

If I had to read the tealeaves, I think it will occur first somewhere outside the USA, due to the influence of Wall Street on your regulators and interference from entrenched interests. Somewhere in Asia first I bet, then Europe will be close behind but just pipped to the post by somewhere in the USA as some west coast state sees sense supporting the locally made product....

I'd put money it'll be somewhere in America first.

 

[renim]

FWIW, hobbyists can now buy real LIDAR at same price as GPU cards....
LeddarTech launches LeddarVu affordable, small LiDAR platform
Livox Mid-40 - DJI Mobile Online Store (United States)

the idea that LIDAR is not needed is fine, the idea that LIDAR won't be used is bizarre.

LIDAR is simply becoming cheaper than GPU. A cheap GPU + cheap LIDAR will be a more resilient combination than an expensive GPU, and also cheaper than an expensive GPU. Elon Musk saying LIDAR is not the future is like Akio Toyoda saying EVs is not the future.

IT history - famous apocryphal sayings
"I think there is a world market for maybe five computers,"
"640kb ought to be enough for anybody"
but software companies made real bets against hardware, like wordprocessing software optimizing for floppy drive use just before hard drives became common.

Compact, Less-Expensive LiDAR Could Lead to Smaller, Cheaper AVs

point is, LIDAR will be ubiquitous, its better to develop with LIDAR because that what they will be deployed with.

 

[EVMeister]

Tesla don't use a GPU-based approach. Just because some form of LIDAR may be cheap doesn't mean the LIDAR required here would be nor that it's right for this task. I know quite a few experts through my work who work with LIDAR and they absolutely agree it's a great tool and works well for its intended purpose. But it is not necessary, helpul nor sufficient for FSD.

 

[Cosmacelf]

So an $800 LIDAR compared to a $2 camera and a maybe $10 AI chip (marginal cost of chip manufacture). And LIDAR doesn't work in heavy rain and snow.

There really is no reason why computer vision can't be solved eliminating the need for LIDAR, especially when you have RADAR backup.