FSD Beta 10.13

[drtimhill]

True FSD needs to be 100% perfect if it is a pathway toward Robo-Taxi.

Actually 90% is indeed on a path to 100%.

And if you always disengage before FSD beta gets to a known problem area its not going to help beta develop much. I'm not talking about disengaging at a spot where you know it will do something dangerous, just something it has trouble coping with (assuming you are sending in reports, that is).

 

[jabloomf1230]

The problem with FSD beta is that 90% is nowhere good enough. True FSD needs to be 100% perfect if it is a pathway toward Robo-Taxi. If there is one serious accident in 100 million miles of FSD, the regulators will seek to outlaw it's use in Level 3 + so Tesla needs to achieve the perfection of a 4 function calculator to be considered ready for wide spread FSD and Robotaxi. This will be near impossible as an FSD Tesla will need to predict the erratic actions of human drivers, especially those humans who intentionally break the law, speeding, DUI, running red lights and stop signs erratic lane changing and road rage. Hopefully, FSD will one day predict illegal driving and using it's AI avoid them with better than good human driver action in advance.

I continue to have great success with FSD beta but I always shut it down when approaching a condition I know FSD-beta fails often. 10% failure is too often in my book.

90% success on Chuck's ULT probably indicates a much higher rate of success on both other ULTs and maneuvers of similar difficulty. My theory about the focus on Chuck's ULT is that Elon seized on it as a way of challenging his AI team. If it can't be done reliably with existing hardware, Tesla may decide to move on to a new hardware setup for future vehicles. I would guess it would consist of adding high res radar in the front bumper.

 

[AlanSubie4Life]

I would guess it would consist of adding high res radar in the front bumper.

Which would be useless with this turn.

The current sensor suite is sufficient for this turn. A human could make the go/no-go decision easily with the existing sensors. The sensors could definitely be better and it would likely help (eventually, in general) though. But I doubt that is currently the limiting factor for this turn, since the cameras would work for a human.

90% success on Chuck's ULT

This would be a terrible result, though a substantial improvement.

Mastering this turn, or any of the many very common similar turns, requires large improvements in perception, path prediction and planning, which FSD currently lacks. So it’s good to focus on one thing to minimize the variables and just work on the basics. I doubt they are intentionally overfitting or anything like that. Should be a lot of carry over if they have a 2x perception range improvement with something like 10x improvement in sensitivity over current sensitivity at the current range. Hopefully this would result in effectively zero error rate at ranges of 100m which is where it is critical.

 

[drtimhill]

Mastering this turn, or any of the many very common similar turns, requires large improvements in perception, path prediction and planning, which FSD currently lacks. So it’s good to focus on one thing to minimize the variables and just work on the basics. I doubt they are intentionally overfitting or anything like that. Should be a lot of carry over if they have a 2x perception range improvement with something like 10x improvement in sensitivity over current sensitivity at the current range. Hopefully this would result in effectively zero error rate at ranges of 100m which is where it is critical.

What do you base all these numbers on? Can you quote any comparative sources? Do you know what Tesla currently achieves, and how that compares to other systems? What does "effectively zero error rate" mean? Why is 100m "critical" (and what do you mean by that?) and under what conditions?

 

[AlanSubie4Life]

What do you base all these numbers on? Can you quote any comparative sources? Do you know what Tesla currently achieves, and how that compares to other systems? What does "effectively zero error rate" mean? Why is 100m "critical" (and what do you mean by that?) and under what conditions?

They’re rough guesses of course. It’s very hard to know what Tesla currently achieves; we can only observe the results and guess at what might be the issue. Effectively zero error rate would ensure low enough error rate that turning in front of an unobscured moving car at critical distance just does not occur, in a practical sense, I guess.

And of course we know roughly what the requirements are, based on time to do maneuvers and the speed of vehicles.

 

[drtimhill]

They’re rough guesses of course. It’s very hard to know what Tesla currently achieves; we can only observe the results and guess at what might be the issue. Effectively zero error rate would ensure low enough error rate that turning in front of an unobscured moving car at critical distance just does not occur, in a practical sense, I guess.

And of course we know roughly what the requirements are, based on time to do maneuvers and the speed of vehicles.

So basically all your "they need XX" etc are just made up numbers with no factual basis. Hmmm.

 

[AlanSubie4Life]

So basically all your "they need XX" etc are just made up numbers with no factual basis. Hmmm.

No. That is incorrect. See above.

I mean, we definitely have some idea of what the requirements are. And we see examples of the car going (incorrectly) in videos when cars are visible. So presumably that is a perception issue. (I hope so anyway!). I mean, I definitely could be wrong, but I definitely am not making things up.

 

[Jeff N]

Waymo and Cruise vehicles crash all the time and aren’t outlawed. Cruise just had a serious collision a few months ago (though it is being investigated, they haven’t been shut down).
Obviously it does need to be very close to 100% on any given maneuver to exceed human reliability.

I’m not aware of Cruise (or even Waymo) vehicles crashing “all the time”. That serious collision with the Cruise vehicle resulted in minor injuries and seems to be the fault of the other car which was driving 40 mph in a 20 mph zone and drove straight through the intersection while in a right-turn-only lane this hitting the Cruise which was making a legal left turn.

 

[JHCCAZ]

...I doubt they are intentionally overfitting or anything like that. Should be a lot of carry over...

To me this is a big question, and I'm not as confident that solving (well let's say "vastly improving performance on") one or two of these famous turns will necessarily carry over to many other similar turns.

It seems that the nature of the NN training is very dependent on the training set, and other scenarios that we would think of as "similar" may not seem so to the NN. This is why the release notes always talk about modest X% improvements based on huge numbers of video clips being applied in training.

Having said that, I would agree that vast Improvement on Chuck's most famous turn would still be an extremely encouraging development, because it would validate the capability of the FSD perception and planning to handle such scenarios. Then, getting to the carryover generalization may still require enormous amounts of video-clip training and/or further NN network architectural tweaking.

So, in ascending order of accomplishment, I see the following possible improvements:

• Significant Improvement in success rate for this turn, as advertised, even if not nearly 100%
  • And as I've said repeatedly, success could include a sensible reroute, eg turn right and U-turn at the light - but there's no evidence they're working on that at all
• Indeed near-100% success rate on this particular turn, like the daily human users.
• Demonstrable significant Improvement for similar turns elsewhere in the area.
• Demonstrable improvement for similar turns in most areas, including significantly different-looking environments.
• Essentially nailing all ULTs into high speed cross traffic, from secondary streets.

In this ascending-accomplishment list, I didn't bother to include other scenarios like ULTs from high-speed traffic onto secondary roads; that would be fantastic but is clearly beyond the scope of what's been advertised for this release.

Nonetheless, in my view accomplishing even the first bullet would be very, very encouraging. I do think, and I think Chuck does too, that the B pillar camera placement is a handicap, even setting aside all other criticisms of modest-resolution video Vision.

 

[Daniel in SD]

I’m not aware of Cruise (or even Waymo) vehicles crashing “all the time”. That serious collision with the Cruise vehicle resulted in minor injuries and seems to be the fault of the other car which was driving 40 mph in a 20 mph zone and drove straight through the intersection while in a right-turn-only lane this hitting the Cruise which was making a legal left turn.

Carefully reading the description of the collision I do not think it would have occurred had the Cruise vehicle had been driven by a human (I don't think fault is relevant when evaluating AV collision rates). We'll have to wait to see the whole report. Cruise and Waymo haven't driven enough miles to assess their rate of severe collisions. Waymo had 6 collisions last month, Cruise had 3. To me that is "all the time" though probably not more than expected driving the number of miles they do in San Francisco and the fact that they report every collision, no matter how minor.

 

[Jeff N]

Carefully reading the description of the collision I do not think it would have occurred had the Cruise vehicle had been driven by a human (I don't think fault is relevant when evaluating AV collision rates).

I haven’t seen any description of the incident which would lead me to think it would not have occurred if the vehicle had been driven by a human.

One could certainly imagine a sequence of events (not reflected in the record) that could lead to that opinion but I haven’t actually seen anything written down by a witness or evidence described in an official report that would reasonably lead to that conclusion. Sure, the possibility is open but the probability — I haven’t seen it.

What specific description leads you to draw that conclusion?

 

[Daniel in SD]

I haven’t seen any description of the incident which would lead me to think it would not have occurred if the vehicle had been driven by a human.

One could certainly imagine a sequence of events (not reflected in the record) that could lead to that opinion but I haven’t actually seen anything written down by a witness or evidence described in an official report that would reasonably lead to that conclusion. Sure, the possibility is open but the probability — I haven’t seen it.

What specific description leads you to draw that conclusion?

"The Cruise AV came to a stop before fully completing its turn onto Spruce Street due to the oncoming Toyota Prius"
They stopped in front of an oncoming car while making a left turn. What human would do that? In my initial read of the media reports I also figured the Prius was 100% at fault.

https://www.dmv.ca.gov/portal/file/cruise_060322-pdf/

 

[drtimhill]

No. That is incorrect. See above.

I mean, we definitely have some idea of what the requirements are. And we see examples of the car going (incorrectly) in videos when cars are visible. So presumably that is a perception issue. (I hope so anyway!). I mean, I definitely could be wrong, but I definitely am not making things up.

I agree, we see the car making mistakes. But trying to quantify stuff like "It needs to be XXX better" when no-one really has any way to do such a quantization is dubious .. there is a tendency for people to think someone who provides quantities has more knowledge than someone who does not, and while in many cases that is true, it can also be used by the cynical to try to create an argument that appears to be more authoritative that it actually is.

I'll give an amusing but real-world example. in 1856 after YEARS of work, including the death of two expedition members, a British team succeed in measuring the height of Mount Everest. It turned out to be exactly 29,000 feet, but the team published the result as 29,002 feet because they felt people would think 29,000 sounded too much like a rough approximation.

 

[Daniel in SD]

I agree, we see the car making mistakes. But trying to quantify stuff like "It needs to be XXX better" when no-one really has any way to do such a quantization is dubious .. there is a tendency for people to think someone who provides quantities has more knowledge than someone who does not, and while in many cases that is true, it can also be used by the cynical to try to create an argument that appears to be more authoritative that it actually is.

I'll give an amusing but real-world example. in 1856 after YEARS of work, including the death of two expedition members, a British team succeed in measuring the height of Mount Everest. It turned out to be exactly 29,000 feet, but the team published the result as 29,002 feet because they felt people would think 29,000 sounded too much like a rough approximation.

They should have used confidence intervals since 29,002 is also wrong. I vote that we quantify things but as a range so as to express uncertainty.

 

[Jeff N]

"The Cruise AV came to a stop before fully completing its turn onto Spruce Street due to the oncoming Toyota Prius"
They stopped in front of an oncoming car while making a left turn. What human would do that? In my initial read of the media reports I also figured the Prius was 100% at fault.

https://www.dmv.ca.gov/portal/file/cruise_060322-pdf/

Thanks! I had not seen the actual (redacted) DMV report before.

 

[JHCCAZ]

(I don't think fault is relevant when evaluating AV collision rates).

I don't see why it wouldn't be. If your position is that we should look at overall traffic safety and collision rates, independent of other details like type of collision, extent of injuries, or fault assignment, well I can appreciate the merit of simplicity. And I agree with your point that a driver can cause or greatly contribute to the cause of an accident, even while being legally cleared.

However, the problem is when we try to figure out just how good is some very good future AV. Let's assume that 50% - no let's even make it 75% - of accidents are avoidable by very good defensive driving behavior from other drivers. That still leaves 25% of accidents being unavoidable.

So if you introduce one perfectly-driving AV into this world, you can statistically eliminate only that 75% of all accidents that we deemed to be avoidable. The AV will likely go longer than a human driver between accidents, but only four times longer. Roughly 2 million miles MTBF instead of 500,000 miles. And by refusing to consider the fault determination, we cannot make the case that the AV is any better than 4X human safety - even if it is engineered perfectly.

While this sounds at first like an encouraging Improvement, it's far from a highly compelling case. Consider that shoulder belts and collision survival engineering have made far more than a 4X Improvement in fatality and serious injury rates, yet it took decades for us to get where we are now regarding public acceptance and industry-wide best practices.

Also, you can be sure that AV detractors would tirelessly highlight those accidents involving the robots, especially if the published statistics routinely ignored the at-fault details.

It's only by including the at-fault metric that we would really capture the (nearly) zero accident rate of our hypothetically perfectly-engineered AVs. So I would say it's not only relevant, but actually essential that the at-fault metric should be the main one.

While I agree that at-fault is not a correct or satisfying binary determination in many cases, I'm saying it can't be ignored.

 

[Daniel in SD]

I don't see why it wouldn't be. If your position is that we should look at overall traffic safety and collision rates, independent of other details like type of collision, extent of injuries, or fault assignment, well I can appreciate the merit of simplicity. And I agree with your point that a driver can cause or greatly contribute to the cause of an accident, even while being legally cleared.

However, the problem is when we try to figure out just how good is some very good future AV. Let's assume that 50% - no let's even make it 75% - of accidents are avoidable by very good defensive driving behavior from other drivers. That still leaves 25% of accidents being unavoidable.

So if you introduce one perfectly-driving AV into this world, you can statistically eliminate only that 75% of all accidents that we deemed to be avoidable. The AV will likely go longer than a human driver between accidents, but only four times longer. Roughly 2 million miles MTBF instead of 500,000 miles. And by refusing to consider the fault determination, we cannot make the case that the AV is any better than 4X human safety - even if it is engineered perfectly.

While this sounds at first like an encouraging Improvement, it's far from a highly compelling case. Consider that shoulder belts and collision survival engineering have made far more than a 4X Improvement in fatality and serious injury rates, yet it took decades for us to get where we are now regarding public acceptance and industry-wide best practices.

Also, you can be sure that AV detractors would tirelessly highlight those accidents involving the robots, especially if the published statistics routinely ignored the at-fault details.

It's only by including the at-fault metric that we would really capture the (nearly) zero accident rate of our hypothetically perfectly-engineered AVs. So I would say it's not only relevant, but actually essential that the at-fault metric should be the main one.

While I agree that at-fault is not a correct or satisfying binary determination in many cases, I'm saying it can't be ignored.

That’s why I’m always saying 10x human safety is impossible. If I’m riding in a robotaxi I want a lower chance of getting injured or dying. Fault is 100% irrelevant to me.
If you make the at-fault metric the main one then a robotaxi could actually make the roads less safe. I’m not saying robotaxi companies should be liable for not at fault collisions just that it is not acceptable for them to have more collisions than human drivers.

 

[JHCCAZ]

...If you make the at-fault metric the main one then a robotaxi could actually make the roads less safe. ...

That's only if the Robotaxis are running around causing many accidents that are technically not their fault, but also wouldn't have occurred with human drivers.

I understand this possibility and that you're assigning that Cruise accident to be such a case. But I don't think it's a realistic concern of well-developed AVs, maybe not even a statistically important concern of less well-developed AVs. Even if we stipulate that the Cruise accident is exactly what you say, and that more of these will occur as things develop, statistics are vastly different than an overemphasis on a few media- or Internet-hyped examples.

If one were looking for reasons not to ride in a Robtaxi, or more generally to oppose AVS on the road, one would start by refusing to consider the details of a large statistical sampling of accident classifications and associated fault determination.

 

[drtimhill]

They should have used confidence intervals since 29,002 is also wrong. I vote that we quantify things but as a range so as to express uncertainty.

The public doesnt understand such things as confidence intervals. As for quantifying .. I'm dubious that there are any measures that make much sense at the moment, given that we are slashing our way through an unexplored jungle with only a compass to help us.

A lot of people have chimed in with stuff like "you need high resolution cameras" ... "you need MORE cameras" .. "You need xxx% accuracy" .. "you need YYY". Realistically, no-one knows at present. Really. We can all speculate and guess, but that's all it is. And that's fun and fine, we all like doing it, as long as everyone realizes that it IS all idle speculation.

Right now it's like the Wright Brothers and flying .. its more guess-work and good luck rather than a defined engineering discipline. Sure, we can test and make sure what we come up with works, but there is no way up-front to know what WILL work except trial and error combined with inspiration (and perspiration of course).

 

[AlanSubie4Life]

It needs to be XXX better"

It’s true we can’t see everything in the visualizations. But when cars flicker in and out at around 100m range, it seems reasonable that that is the current limit of the perception. Hopefully it is actually better than that but the actual behavior of the car suggests otherwise.

Since that’s about 4 seconds, and that’s not enough time to make the maneuver without freaking people out, clearly we need at least double that, so 200m or so reliably (obviously depends on the speed of both oncoming vehicles and the speed of maneuvering of the FSD vehicle).

That’s where the 2x number I guesstimated came from. I don’t think it is too far off (in this case), though I’m open to other thoughts on this.

And the 10x better sensitivity is definitely more of a WAG. We see a lot of flickering at 100m and we’d like to see (a lot) less of that at 200m.

And it may not work this way, but my assumption is that probability of detection drops off quite quickly at increasing range and so closer in, if we are doing well at 200m, should be a lot better than now at 100m (but this is definitely less clear and there may be a “floor” on the close-in error rate (this would be bad)).

Anyway as you said I don’t really know how it works. But the TLDR is that I think it needs substantially better performance based on what we have seen, and it is possible to sort of quantify what that would look like, and I think it is hard to argue that it is good enough.

I don’t see how they’ll be able to comfortably do this left turn reliably without at least a part of that step change in capabilities.