Automatic driving on city streets coming later this year!

[daniel]

The International Telecommunications Union defined mobile communications, lte, 4g, 5G etc. It didn’t work out as planned. To the best of my knowledge, Tesla has never “defined” their autopilot to be any specific level. They talk about “features”, but don’t get SAE specific. I bet that is a planned strategy. Tesla likely will not define their FSD level until it “passes” regulation in Europe, USA and China. I expect them to be “feature complete” with a “driver” in the seat... I will probably be happy with that.

The development of robo-taxis and Tesla network might take years, “feature complete” could be 2020. SAE certification will be applied after regulations are “passed” and accepted likely years from now.

Tesla was very specific that 2018 Teslas, if the buyer paid for FSD, would be capable of driverless operation anywhere a human could drive. They may not have used the term L5, but that is what they were promising. They also declared absolutely that those (HW2.5) cars had all the necessary hardware and were only waiting for the software and permitting. Now they've backed away from that, and driverless cars are just a vague goal for some nebulous time in the future. And now that it turns out that the HW3 computer is needed just for the next baby step past EAP they're struggling to try to figure out how to upgrade the HW2.5 cars because right now it's a very difficult and time consuming operation.

 

[Daniel in SD]

The International Telecommunications Union defined mobile communications, lte, 4g, 5G etc. It didn’t work out as planned. To the best of my knowledge, Tesla has never “defined” their autopilot to be any specific level. They talk about “features”, but don’t get SAE specific. I bet that is a planned strategy. Tesla likely will not define their FSD level until it “passes” regulation in Europe, USA and China. I expect them to be “feature complete” with a “driver” in the seat... I will probably be happy with that.

The development of robo-taxis and Tesla network might take years, “feature complete” could be 2020. SAE certification will be applied after regulations are “passed” and accepted likely years from now.

Tesla could operate robotaxis in California today, the regulations are already passed.
Tesla has defined their vehicles as capable of Level 5 autonomy.

Basic news is that all cars exiting the factory have hardware necessary for Level 5 Autonomy so that’s in terms of Cameras, Compute Power, it’s in every car we make on the order 2,000 cars a week are shipping now with Level 5 literally meaning hardware capable of full self-driving for driver-less capability.

I agree that they have walked that back for new orders but FSD v1 was definitely promised to be Level 5.

 

[Thp3]

Even without Tesla giving us levels, we can figure out the levels on our own since we know the SAE definitions. For example, we know that Autopilot is currently L2 even without Tesla telling us. Similar with FSD, Tesla does not need to tell us the levels for us to know what level it is. When Tesla removes the nags and removes the requirement of the driver to pay attention, that will be a pretty big clue that AP has become an autonomous system, L3 or above. And, at that point, it will be easy to figure out if it is L3, L4 or L5 too. If the ODD is limited, we will know that it is L3 or L4 depending on whether the car can handle its own fallback and if the ODD is not limited, we will know that it is L5, because that is what the SAE says.

I suspect level definitions will be fuzzy. Tesla will keep the “nags” as long as regulations require a driver in the “driving seat”. It doesn’t make any difference what HW3 can do if regulations define a steering wheel, rear view mirrors, etc. The nags might decline or be easier to defeat, but existing car regulations will limit the clear use of SAE levels to compare selfdriving software from car manufacturers. All will be limited by regulations from car accidents in the last 50 years.

I don’t think we will benefit from the categorization except in the ideal world where existing regulations don’t exist.

Our understanding and education will improve by defining “edge cases” for the “feature complete” autopilot. Teaching each other the edge cases to be wary of, will keep us safer. I’m quite vigilant on highways that allow cross traffic when using autopilot. The worst ones are the near-freeways with cross streets and left turns.

I do wonder what the “edge cases” in city driving will be? Dogs? Kids darting out chasing balls? Left turns at five-way intersections? I expect to be surprised.

 

[Thp3]

Tesla could operate robotaxis in California today, the regulations are already passed.
Tesla has defined their vehicles as capable of Level 5 autonomy.

I agree that they have walked that back for new orders but FSD v1 was definitely promised to be Level 5.

On paper in California, Tesla could operate if they filed the paperwork and gave the state government access to some of the data. So far, I think Tesla has declined to “pilot” any selfdriving cars in California.

It will be interesting to see how Tesla proceeds.

I expect city driving with NOA and some limitations will be quite close to FSD for many of us in 2020. Having the car drive us to “work” while we read email in the back seats, drop us off and return 8 hours later to take us home probably won’t happen in 2020. But it’s something to wish for...

 

[Daniel in SD]

On paper in California, Tesla could operate if they filed the paperwork and gave the state government access to some of the data. So far, I think Tesla has declined to “pilot” any selfdriving cars in California.

Tesla does have a permit to test AVs in California but I don't think they've done so since 2016. That's how they created the infamous "FSD" demo video. It will be interesting to see if they reported the testing they did in 2019 (the autonomy day test drives) or if they're going to try to go the Uber route and argue that their vehicles are not autonomous (that didn't work out for Uber).
The only thing fuzzy about Tesla's level definitions is whether they will be able to continue to sell FSD as a Level 2 system or if the DMV will determine that FSD is actually a prototype Level 3-5 system and require them to comply with testing regulations. My opinion is that automatic driving on city streets as demoed on Autonomy day is a prototype AV and will be regulated as such (others disagree! I'm hoping that we find out who was right this year. Unfortunately the more likely scenario is that automatic driving on city streets will not be released this year).

 

[daniel]

Tesla could operate robotaxis in California today, the regulations are already passed.
Tesla has defined their vehicles as capable of Level 5 autonomy.

Basic news is that all cars exiting the factory have hardware necessary for Level 5 Autonomy so that’s in terms of Cameras, Compute Power, it’s in every car we make on the order 2,000 cars a week are shipping now with Level 5 literally meaning hardware capable of full self-driving for driver-less capability.

They have permission. But the quote only says that the cars have all the hardware for Level 5. They are far from having the software. And they can't run robotaxis without the software. Note that the above statement was made about my HW2.5 car as well. Tesla said my car has all the needed hardware for robotaxi operation. I didn't believe it then, and the requirement for HW3 just to get past EAP shows that I was right. And I don't believe it now either. Tesla will not operate robotaxis until it can do so safely. Maybe in five years in very narrow geographical areas with perfect streets.

My prediction: In three years they'll have a "feature complete" City NoA operating at Level 2.

 

[Daniel in SD]

But the quote only says that the cars have all the hardware for Level 5. They are far from having the software

Yeah, that's the genius of Tesla's marketing! There is probably no way to prove that even HW2.0 is not capable of Level 5 autonomy with the right software.

 

[jebinc]

Come on, you know what I meant! Yes, I expressed myself poorly. But what I was saying was clear: Tesla was saying that if I paid for FSD my 2018 car would be capable of driverless operation. Now they're not even promising L3 operation to people who pay for FSD today. They're just saying that FSD is their ultimate goal. All they're promising to people who pay for FSD today is City NoA at Level 2. Not even L3. They're not even promising that AP on the highway will become L3, which would be far less than the L5 everywhere they were promising on the cars being sold in 2018.

Tesla has backed way off from the promises it was making, which is a clear tacit admission that they cannot do what they were promising in a time frame that would be of any use for the cars it was being promised on. And at some point they will have to compensate the people who paid for driverless operation, and who, realistically, are not going to get that on their present car.

The hardware is just not there in existing cars, putting aside for the moment the status of their software development.

 

[diplomat33]

Back on the topic of automatic city driving, Tesla is making really nice progress with the visualizations. Tesla just added more road markings, including parking spots for the disabled in the FSD visualizations (2019.40.50.7).

See more here:

This does make me more optimistic for automatic city driving. If the car can really see all these things and have the right driving policy, we might actually get city self-driving.

 

[daniel]

Yeah, that's the genius of Tesla's marketing! There is probably no way to prove that even HW2.0 is not capable of Level 5 autonomy with the right software.

It's seldom possible to prove a negative. But Tesla has admitted that HW2.0 and HW2.5 are inadequate for driverless operation. It's my opinion that HW3 will prove to be inadequate for driverless operation as well, but that's just my opinion, and I won't be too surprised if City NoA becomes available at Level 2 (constant driver supervision and driver responsible for deciding when to intervene) in three years.

L2 City NoA would not be a big enough improvement over EAP to get me to upgrade. But I'm very much looking forward to seeing how L2 City NoA handles edge cases such as one I mentioned before: A bicyclist coming from the opposite direction on the opposite shoulder of the road will cause a car coming up behind the cyclist to swerve slightly into my lane to give the cyclist more room even though that would not be strictly necessary. My car should move to the opposite side of my lane in anticipation, to give that car more room, rather than slavishly keeping the middle of my lane. Will Tesla consider this a situation where the driver is expected to take over? Or will the car accommodate it?

There are other, less complicated situations, when the car should keep to one side or the other of the lane rather than dead-center. When will autosteer become good enough to do this? Right now, with EAP, when I see a vehicle approaching from the other direction that is too close to the center line, I disengage autosteer and move away from the line.

 

[emmz0r]

I am not sure how L2 city NoA would work with sharp turns, as you have to either follow the violent steering wheel turns with your hands, easily disengaging it by accident, or having your hands hovering over it. Does not sound like a useful feature to me.

 

[diplomat33]

I am not sure how L2 city NoA would work with sharp turns, as you have to either follow the violent steering wheel turns with your hands, easily disengaging it by accident, or having your hands hovering over it. Does not sound like a useful feature to me.

I imagine you would just keep your hands near the steering wheel. You won't need to actually hold the wheel during the turn.

 

[jebinc]

Back on the topic of automatic city driving, Tesla is making really nice progress with the visualizations. Tesla just added more road markings, including parking spots for the disabled in the FSD visualizations (2019.40.50.7).

View attachment 495952

See more here:

This does make me more optimistic for automatic city driving. If the car can really see all these things and have the right driving policy, we might actually get city self-driving.

A contrarian thought - As we know, Tesla pumps out OTA firmware updates regularly. Often these updates "break" (a "common practice today") or worsen functionality previously released (e.g., AP as a common example/complaint). For L3 and beyond to become a reality, Tesla would either need to demonstrate to regulatory agencies that they have the capability for this "common practice today" to never happen at L3 and beyond releases, OR we would see very infrequent Tesla firmware releases, from L3 forward, as the release of each subsequent L3+ software package would need to go through extensive testing regulatory review - A testing ad review process not unlike the FAA testing and review process used to "certify" software on airplanes. Given this, I stand fairly firm that our existing Tesla's, with existing sensors/hardware, will not likely ever get beyond "cool/fun" L2 capability (and "parlor tricks").

 

[diplomat33]

A contrarian thought - As we know, Tesla pumps out OTA firmware updates regularly. Often these updates "break" (a "common practice today") or worsen functionality previously released (e.g., AP as a common example/complaint). For L3 and beyond to become a reality, Tesla would either need to demonstrate to regulatory agencies that they have the capability for this "common practice today" to never happen at L3 and beyond releases, OR we would see very infrequent Tesla firmware releases, from L3 forward, as the release of each subsequent L3+ software package would need to go through extensive testing regulatory review - A testing ad review process not unlike the FAA testing and review process used to "certify" software on airplanes. Given this, I stand fairly firm that our existing Tesla's, with existing sensors/hardware, will not likely ever get beyond "cool/fun" L2 capability (and "parlor tricks").

I imagine that if Tesla achieves some level of autonomous driving that OTA updates would have to change since you are correct that Tesla could not release updates that "break" autonomous driving. So the updates will probably be less frequent and undergo more rigorous testing. Also, remember that once Tesla achieves autonomous driving, there shouldn't be any major features left to implement. At that point, it should just be a matter of tweaking the software. So there should less need for big updates that could break the software.

 

[jebinc]

I imagine that if Tesla achieves some level of autonomous driving that OTA updates would have to change since you are correct that Tesla could not release updates that "break" autonomous driving. So the updates will probably be less frequent and undergo more rigorous testing. Also, remember that once Tesla achieves autonomous driving, there shouldn't be any major features left to implement. At that point, it should just be a matter of tweaking the software. So there should less need for big updates that could break the software.

But, doesn't your logic/assumption set require that Tesla move, in "Big Bang" fashion, from some level of L2 functionality to full L3+ level of autonomy? It seems their approach is more of an "incremental" implementation of features, rather than a "full up" big bang release for L3, L4, or L5, no?

 

[diplomat33]

But, doesn't your logic/assumption set require that Tesla move, in "Big Bang" fashion, from some level of L2 functionality to full L3+ level of autonomy? It seems their approach is more of an "incremental" implementation of features, rather than a "full up" big bang release for L3, L4, or L5, no?

Technically, a system just needs a complete OEDR within its ODD to qualify as a L3, L4 or L5 autonomous prototype. L2 is defined as a system with an incomplete OEDR within its ODD so as Tesla releases features incrementally, once they reach a complete OEDR, then technically the system would automatically jump from L2 to a L3+ prototype. Of course, achieving sufficient safety and reliability in your prototype where you can remove driver supervision is another issue.

It is entirely possible that Tesla can achieve a L3, L4 or L5 autonomous prototype with safety driver, on the current hardware and sensors since to do that, they technically just need to finish their vision NN and finish the OEDR. But the reason, I say that the current hardware is not good enough for true autonomy, is because I don't think Tesla can achieve the level of safety and reliability necessary to ever remove the safety driver, certainly not for L5 at least. So while the current hardware might conceivably be good enough for an autonomous prototype with safety driver, I don't think it is good enough for "sleep in the back seat" autonomy.

 

[jebinc]

Technically, a system just needs a complete OEDR within its ODD to qualify as a L3, L4 or L5 autonomous prototype. L2 is defined as a system with an incomplete OEDR within its ODD so as Tesla releases features incrementally, once they reach a complete OEDR, then technically the system would automatically jump from L2 to a L3+ prototype. Of course, achieving sufficient safety and reliability in your prototype where you can remove driver supervision is another issue.

It is entirely possible that Tesla can achieve a L3, L4 or L5 autonomous prototype with safety driver, on the current hardware and sensors since to do that, they technically just need to finish their vision NN and finish the OEDR. But the reason, I say that the current hardware is not good enough for true autonomy, is because I don't think Tesla can achieve the level of safety and reliability necessary to ever remove the safety driver, certainly not for L5 at least. So while the current hardware might conceivably be good enough for an autonomous prototype with safety driver, I don't think it is good enough for "sleep in the back seat" autonomy.

What ODD do you think Tesla could get approval to call L3 (or higher) with existing sensors, single front radar, and cameras (assuming HW3 computer)? Smart Summon, geofenced to a 200' parking lot?

 

[diplomat33]

What ODD do you think Tesla could get approval to call L3 (or higher) with existing sensors, single front radar, and cameras (assuming HW3 computer)? Smart Summon, geofenced to a 200' parking lot?

As I mentioned before, probably something like limited access, divided highways, in clear weather, outside of city limits, clear lane lines, no construction zone, and with light traffic. Under all those conditions, the car would just need to drive on a mostly straight well marked highway in good visibility with no obstructions and with little traffic and no cross traffic, so the current hardware should be able to handle that.

 

[DanCar]

... with light traffic. ....

VW and Tesla have said that heavy traffic, bumper to bumper traffic is easier. Lower speed makes things less dangerous, and just having to follow the car in front is easy.

 

[diplomat33]

VW and Tesla have said that heavy traffic, bumper to bumper traffic is easier. Lower speed makes things less dangerous, and just having to follow the car in front is easy.

True. Tesla could probably do L3 highway in bumper to bumper traffic.

I was thinking of moderate traffic where cars might cut in aggressively or a car is moving very fast in your blind spot. Those would be iffy situations that the current hardware and sensors might not able to totally handle 100% without driver supervision. So I was thinking of traffic conditions that would exclude those cases.