Tesla autopilot HW3

[Daniel in SD]

This isn't like EVs, where the only real downside is range per dollar, and people who are legitimately affected by the downside (as opposed to just being scared of it) are rare. There are a lot of downsides to shared taxis, even if you don't factor in questions about self-driving safety, liability, etc., with cost being the biggest.

People driving more than 300 miles in a day are at least a couple of orders of magnitude less common than people traveling for more than half an hour each way. Even with a 30-minute commute, shared cars start to be problematic in terms of electricity spent, because the odds of the car having a passenger to take during its the 15-to-20-minute reverse commute are low. So you can assume that about 30% of the driving will be just moving the car from one place to another without anybody in the car. That adds to the cost significantly.

Also, self-driving taxis will require lots of extra maintenance, or else they will have all the same problems with cleanliness as taxis and city buses, only worse. People abuse things that they don't own, and if you think that your average taxi is in bad shape, can you imagine if it had no driver? Without cleaning them at least once a day (and maybe more), these things will smell like pee within a month, if not a week. Nobody wants to endure that for even a half-hour commute.

All of those extra costs add up. If taxi costs are any indication, the cost per mile of self-driving taxis, even if you take the cost of the driver out of the question, is likely to be about twice as expensive per mile as owning a personal vehicle, even after factoring in the cost of the vehicle (assuming your car lasts 200k miles). There's just no way that self-driving taxis can be nearly as cheap as driving a car that you own, in practice, unless you're one of those people who buys a new car every three years just to have a new car (and if you are, then you probably won't be caught dead riding in a car that smells like vomit because of the guy it took home from a bar the night before).

The folks who think self-driving taxis will replace personal vehicles clearly haven't done the math. The economics just don't work unless you live in one of those horrible places where you have to pay an extra $300 a month to rent a parking space.

I wish I could read French! Anyway, he posts this table:

It seems like the only costs here that would apply to self driving vehicles are fuel, insurance, maintenance and lease payments. Insurance will be far cheaper since self driving cars, if they ever exist, will be far safer than humans. Fuel could theoretically be much cheaper since self driving cars can be mostly short range EVs, there's no need to have extra battery capacity when someone is only ordering a 30 mile ride. I'm not sure that keeping the cars clean will be a huge problem. I'm guessing it will be like Uber and Lyft where they'll fine you if you pee in the car. I think traditional cabs and public transit are not a good comparison since they have no way to ban or punish abusive users.

 

[J1mbo]

You may be right. There also might be more to the system than we were ever told.

One of the interesting things I found in the public electronic parts catalog that I've never known quite what to make of is what appears to be a rear radar as part of the AP2.5 suite. Tesla has never said a word about it that I've seen - but look at this:

https://epc.tesla.com/#/systemGroups/66322

View attachment 468670

Assembly 7 appears to be a rear mounted radar unit... I haven't taken the bumper off to see if it is physically installed on my Raven, and the EPC doesn't spell out an exact location. Also don't know what to make of the front corner radar brackets mentioned here.

It might be interesting to take a battery powered radar detector around a modern Tesla and see how many radars it actually has and where they cover...

That is interesting. We know that the MX was originally going to launch with an updated Mobileye suite, which was pulled at the last minute, but some parts made their way to production cars (such as the dual camera housing with only one camera in it) - but this radar is for HW2.5 cars. I am sure it would have been fully dissected on here if it was actually fitted in the wild!

 

[bdub85]

This isn't like EVs, where the only real downside is range per dollar, and people who are legitimately affected by the downside (as opposed to just being scared of it) are rare. There are a lot of downsides to shared taxis, even if you don't factor in questions about self-driving safety, liability, etc., with cost being the biggest.

People driving more than 300 miles in a day are at least a couple of orders of magnitude less common than people traveling for more than half an hour each way. Even with a 30-minute commute, shared cars start to be problematic in terms of electricity spent, because the odds of the car having a passenger to take during its the 15-to-20-minute reverse commute are low. So you can assume that about 30% of the driving will be just moving the car from one place to another without anybody in the car. That adds to the cost significantly.

Also, self-driving taxis will require lots of extra maintenance, or else they will have all the same problems with cleanliness as taxis and city buses, only worse. People abuse things that they don't own, and if you think that your average taxi is in bad shape, can you imagine if it had no driver? Without cleaning them at least once a day (and maybe more), these things will smell like pee within a month, if not a week. Nobody wants to endure that for even a half-hour commute.

All of those extra costs add up. If taxi costs are any indication, the cost per mile of self-driving taxis, even if you take the cost of the driver out of the question, is likely to be about twice as expensive per mile as owning a personal vehicle, even after factoring in the cost of the vehicle (assuming your car lasts 200k miles). There's just no way that self-driving taxis can be nearly as cheap as driving a car that you own, in practice, unless you're one of those people who buys a new car every three years just to have a new car (and if you are, then you probably won't be caught dead riding in a car that smells like vomit because of the guy it took home from a bar the night before).

The folks who think self-driving taxis will replace personal vehicles clearly haven't done the math. The economics just don't work unless you live in one of those horrible places where you have to pay an extra $300 a month to rent a parking space.

i see them as 2nd or 3rd car which is basically an investment or business. I can’t see using my main vehicle as an shared taxi. The only way I see it working is for people who barely drive their cars but if that’s the case why aren’t they just using a shared taxi?

 

[J1mbo]

i see them as 2nd or 3rd car which is basically an investment or business. I can’t see using my main vehicle as an shared taxi. The only way I see it working is for people who barely drive their cars but if that’s the case why aren’t they just using a shared taxi?

This will be a shared taxi - but cheaper, because there will be no driver to pay, and the car will be so safe that insurance will be essentially free /s

 

[Pale_Rider]

You may be right. There also might be more to the system than we were ever told.

One of the interesting things I found in the public electronic parts catalog that I've never known quite what to make of is what appears to be a rear radar as part of the AP2.5 suite. Tesla has never said a word about it that I've seen - but look at this:

https://epc.tesla.com/#/systemGroups/66322

View attachment 468670

Assembly 7 appears to be a rear mounted radar unit... I haven't taken the bumper off to see if it is physically installed on my Raven, and the EPC doesn't spell out an exact location. Also don't know what to make of the front corner radar brackets mentioned here.

It might be interesting to take a battery powered radar detector around a modern Tesla and see how many radars it actually has and where they cover...

Hmm, if you google 1128429-00-D, you get a bunch of front radar brackets. I wonder which is correct, the parts catalog (you'd think) or all these parts out there that may have the part number on them...

 

[Pale_Rider]

You may be right. There also might be more to the system than we were ever told.

One of the interesting things I found in the public electronic parts catalog that I've never known quite what to make of is what appears to be a rear radar as part of the AP2.5 suite. Tesla has never said a word about it that I've seen - but look at this:

https://epc.tesla.com/#/systemGroups/66322

View attachment 468670

Assembly 7 appears to be a rear mounted radar unit... I haven't taken the bumper off to see if it is physically installed on my Raven, and the EPC doesn't spell out an exact location. Also don't know what to make of the front corner radar brackets mentioned here.

It might be interesting to take a battery powered radar detector around a modern Tesla and see how many radars it actually has and where they cover...

Looks like 1108647-00-D is the radar part, which is the same part number as the Model 3 front radar, continental part used in all AP2.5+ cars. The other radar part number is for the Bosch AP1/2.0 cars. If i had to guess, the one diagram is for AP1 (which had some FW references to corner radar I believe) and AP2.0 radar hardware and the top diagram is for AP 2.5+ configurations.

 

[Cirrus MS100D]

Don’t forget that when the Model 3 first came out, everybody thought the box on the back was rear radar, it was actually the Bluetooth radio. Do they potentially use the same bracket (for cost-cutting measures)?

 

[Saghost]

Random musing from a dark rain soaked Autopilot commute: Autopilot has no object permanence.

It seems to me like the current state of the software is only analyzing the current frame and reacting to it - hence the vehicles that switch back and forth between classes, and occasionally appear and disappear, and the lane lines that rapidly jump back and forth when a merge happens.

The car is clearly quite good at analyzing and reacting to the current frame, but it seems to me like there is significant value in connecting those together - in comparing the cars and lines it sees in this frame and their locations to the cars and lines it saw in the last frame.

I would expect that this has the potential to make things smoother, to react earlier to some threats, and to help with cases where conditions obscure some of the necessary data.

 

[mongo]

Random musing from a dark rain soaked Autopilot commute: Autopilot has no object permanence.

It seems to me like the current state of the software is only analyzing the current frame and reacting to it - hence the vehicles that switch back and forth between classes, and occasionally appear and disappear, and the lane lines that rapidly jump back and forth when a merge happens.

The car is clearly quite good at analyzing and reacting to the current frame, but it seems to me like there is significant value in connecting those together - in comparing the cars and lines it sees in this frame and their locations to the cars and lines it saw in the last frame.

I would expect that this has the potential to make things smoother, to react earlier to some threats, and to help with cases where conditions obscure some of the necessary data.

The UI may be fed from a less filtered version of the data stream. The high level control logic smooths path routing and needs to track brake lights vs turn signals so there is some persistence there. But yeah, the low level NN reportedly gets two frames and deduces everything from that.


Regarding object permanence (and knowing the unknowns), I do wonder about situations like turns across multiple lanes where there could be a car in shadow of the visible one that you had previously seen.

 

[Saghost]

The UI may be fed from a less filtered version of the data stream. The high level control logic smooths path routing and needs to track brake lights vs turn signals so there is some persistence there. But yeah, the low level NN reportedly gets two frames and deduces everything from that.


Regarding object permanence (and knowing the unknowns), I do wonder about situations like turns across multiple lanes where there could be a car in shadow of the visible one that you had previously seen.

Is there evidence that the current version of the software is reacting to either brake lights or turn signals?

I haven’t noticed my car behave differently in response to either of those, only the motion of the cars around it.

 

[mongo]

Is there evidence that the current version of the software is reacting to either brake lights or turn signals?

I haven’t noticed my car behave differently in response to either of those, only the motion of the cars around it.

Not sure, I live vicariously through the reports of others.
From the autonomy presentation, I though Karpathy was talking about detecting non-signally lane changers, which implies they detect signalling lane changers.

 

[MarcEMarc]

What I’ve noticed is that the v10 update did improve a lot of the autopilot image functionality with better lane lines and traffic images. A few things I have also noticed:

1) there is significant lag when an image (car/truck/motorcycle) goes from one camera field to another (e.g. moving slowly next to the vehicle). It flickers one then the other position and dances a little bit (not as bad as prior to v10).

2) EVERY time I have been using Autopilot and it starts to rain hard, all the functionality of Autopilot goes away with a warning the lane departure is inactive and canceling the Autosteer. This is a serious and significant downfall of camera based navigation that doesn’t happen with radar or lidar. I really don’t see any future in solving this problem with just image cameras unless there some additional imaging technologies.

 

[emmz0r]

What I’ve noticed is that the v10 update did improve a lot of the autopilot image functionality with better lane lines and traffic images. A few things I have also noticed:

1) there is significant lag when an image (car/truck/motorcycle) goes from one camera field to another (e.g. moving slowly next to the vehicle). It flickers one then the other position and dances a little bit (not as bad as prior to v10).

2) EVERY time I have been using Autopilot and it starts to rain hard, all the functionality of Autopilot goes away with a warning the lane departure is inactive and canceling the Autosteer. This is a serious and significant downfall of camera based navigation that doesn’t happen with radar or lidar. I really don’t see any future in solving this problem with just image cameras unless there some additional imaging technologies.

1) because of (kalman) filtering and latency between the AP computer -> MCU input -> MCU 3D processing and arrangement -> kalman filtering of 3D objects -> rendering

2) LIDAR would be even worse. Every rain drop and fog particle will reflect the laser.

 

[diplomat33]

2) EVERY time I have been using Autopilot and it starts to rain hard, all the functionality of Autopilot goes away with a warning the lane departure is inactive and canceling the Autosteer. This is a serious and significant downfall of camera based navigation that doesn’t happen with radar or lidar. I really don’t see any future in solving this problem with just image cameras unless there some additional imaging technologies.

I've noticed this as well. Presumably, better image processing could at least bring it up to parity with what humans can do. But obviously, weather conditions that even human drivers can't handle, would also remain impossible to the AP hardware.

I do suspect that the current hardware is probably limited to L4 autonomy because it will be limited to only fair weather. L4 autonomy would still be amazing though but obviously not as good as L5.

 

[Pale_Rider]

I’ve never had rain disable autopilot. NoA yes, but not autopilot. I live in Houston and recently had to drive around during a tropical storm... And while radar can see through rain fine, it can’t see lanes or read signs. Lidar really struggles in rain, snow, fog... It would shut down long before vision based solutions...

 

[Leifp]

Hmmm, so what happens if I send my Model 3 out as a Robotaxi and it starts to rain?

 

[jebinc]

Hmmm, so what happens if I send my Model 3 out as a Robotaxi and it starts to rain?

You light a candle and pray he/she comes home at the end of the day? Just having fun!

 

[NOLA_Mike]

Hmmm, so what happens if I send my Model 3 out as a Robotaxi and it starts to rain?

I'm imagining it would pull over, stop, turn on the hazards and resume when it can see again. But I damn sure don't know.

Mike

 

[Saghost]

Hmmm, so what happens if I send my Model 3 out as a Robotaxi and it starts to rain?

If the software then still can’t drive safely, I would expect it to pull over and wait for the weather to clear.

 

[diplomat33]

Hmmm, so what happens if I send my Model 3 out as a Robotaxi and it starts to rain?

Well, presumably a robotaxi would need to be programmed to pull over safely if conditions became too difficult for it to handle.

In my situation, AP can handle mild to medium rain. But it cannot handle a heavy deluge where a human would struggle. The one instance where AP quit on me was in a very heavy downpour like Noah's flood where I could not see anything either and almost pulled over myself.