Search results

Can't use L3 if you're drunk.

Had to get rid of all the CNNs…

Maybe they really are gating the release of V12.4 on Chuck's turn? 10x fewer interventions would be great progress (though probably need 100,000x fewer for Robotaxi.) I wonder how many times in a row Chuck could do it without crashing.

Waymo is liable when their test drivers crash too (same as any other company employing drivers.) I think you can assume the Waymo Driver would have had a collision whenever they say “shortly”. The paper where they looked at all the collisions they had in Phoenix didn’t describe any cases where...

You’re going to look pretty dumb when a FOIA request shows that the NHTSA is requiring FSD to stop 10 feet behind the line.

New vehicle! Hopefully this will fix the bad training data problem. I think Chuck also has a Model Y.

They said the car ran the red light because of human error by remote assistance. Clearly remote assistance changes what the Waymo Driver is allowed to do.

The Zoox robotaxi does not have a front and back. https://zoox.com/vehicle

Yeah that’s obviously not true though. For example we know that remote assistance can run red lights. However I’m not saying they intend to have remote assistance be able to run into a telephone pole. I’m saying I could see remote assistance code being way more likely to have bugs.

Obviously it should still be functioning. Lots of ways to screw up programming. Maybe the transition between autonomous and remote assistance maneuver has a bug. Maybe remote assistance can give the car authority to hit stationary objects in some cases (low tree branch? Tumbleweed? Balloon?)...

Yeah. I’m just saying it’s much more likely for there to be bug in fleet response mode than normal autonomous mode. Figuring out where to stop to pick people up is something Waymo and Cruise have struggled with. It wouldn’t surprise me if it’s one of the most common causes of fleet response.

My guess is the Waymo was using remote assistance and there is a bug that allowed it hit the poll. They’ve driven 7 million driverless miles without any other known collisions with fixed object so it seems unlikely it was in its normal operating mode.

Watching the video of the actual remote interface I don't really see any significant lag. This could just be an artifact of a fun demo mode (obviously you can't actually pause real life).

Thanks, wasn't too bad. Very tired and achy for a few days with low level fever. I assume one these "flirt" variants are what got me (right at the yearly low point for circulation!).

Just got Covid for the first time last week so 0% effective as far as I'm concerned. :p Last year they presented a bunch of data on the previous vaccine when they were approving the new one.

Being charged with homicide is a legal reason not to use it unsupervised... https://electrek.co/2024/04/23/tesla-driver-arrested-homicide-running-over-motorcyclist-autopilot/ Though I do agree that the implication of that advertisement is that FSD was close to complete.

Yes, automation complacency is a big issue with these systems. Looking in the manual I'm surprised they don't have specific warning for large vehicles going across the path of travel. There have been quite a few fatal collisions where people have hit the side of trailers (train cars seem very...

I would strongly suggest not trying a third time. I feel like your case would be stronger if you didn’t say it was the second time. Did you crash the first time too? Anyway this is a well known issue and there is a warning in the manual.

I’ve made your robotaxi design.

I think it's pretty hard to get at (it's between the door and the side of car.) Someone here with a minivan can probably chime in but I expect you'd need a metal rod to get at it and if the vehicle needs to be metal rod resistant from the inside that's going to be pretty tough!

Ah. I think I've noticed the wiring before. Here's a picture of it. It needs to be about half the slide length.

They've been training FSD on this turn for years. Chuck claims to have met the engineer who created the simulation of it. Not sure how left turns by drivers coming from the right changes things. Those drivers have the right of the way.

It’s a simple unprotected left turn. Whether unprotected left turns are simpler than other types of turns is a much more complicated question. Unprotected left turns are simple because it is very clear who has the right of way and you don't have to worry about mistakes made by other drivers...

Why? They roll down on minivans.

Are we talking about the same turn? I can see fine looking at YouTube video. I've never seen Chuck have to abort for something that wasn't easily visible on the wide angle camera (which is far lower resolution than human eyes).

Not betting on 12.4. No way they're going to get it working by next week.

UPL Team is definitely under the gun!

It was brought up because @HitchHiker71 suggested that the Tesla Robotaxi would initially be L3. Remote operation would allow Tesla to operate many robotaxis per driver. The only other application I can think of is delivering a rental car to someone or paying lower wages to drivers by...

Capabilities and Performance Unlike active safety systems, such as automatic emergency braking(AEB), that intervene only momentarily when necessary to help prevent a collision, active driving assistance systems provide continuous support intended to make driving easier—for instance, when you’re...

Does sound like it. I know someone with some extra American made AVs

Lots of stuff is legal in Nevada.

There are companies that claim it’s possible to safely remotely drive cars. https://vay.io/

Why not use remote assistance so you can have a single “driver” run many robotaxis?

To be pedantic you’re responsible for monitoring that a wheel is about to fall off or the car is on fire.

They probably do the stop sign training with simulated data.

Most of their driving is during the day on Los Angeles interstates (where speed is very often under 40 mph). Or they want an L2 system with a top speed of 130mph. Most likely though they just want a Mercedes instead of a Tesla.

Waymo data from Phoenix showed a very different distribution from human drivers. In 6 million miles only 30 of least severe category (real and counterfactual simulation) and 8 with airbag deployment. Also note that there were zero collisions that did not involve another human...

Why not use Tesla’s 1 per million mile metric? Severe collisions are what really matter. Talking about collisions that are equivalent to curb rash just confuses things. The evidence shows you can’t extrapolate from L4 collision rate to severe collision rate.

Are you saying there are collisions Cruise didn’t report? How many L4 collisions did they have? Not even sure what point you’re trying to make by bringing up L4 collisions.

Robotaxis will not have any L4 collisions so they’re not relevant.

Not very useful for robotaxi safety because the distribution might be completely different (it is for Waymo).

I think @Doggydogworld is correct. You can look for notable Cruise and Waymo collisions on the DMV website and they're not there. The data for deployed vehicles is here...

I predicted it would never happen. Nobody will ever release a competitor to FSD (supervised) in the US.

"12.4 has almost completely retrained models. The final touches are for comfort" giving me big regression vibes. I feel that FSD progress should really be the march of 1's, the binary number system is more appropriate for computers and will let us better gauge progress. We've already got two 1's...

Our original bet Jul 15th 2022. Stock price $240. You knew it would be years of failure and yet did not sell. Sad! End-to-end will get the first nine and send stock to the moon. Bet is still on for 12.5?

They lay off the supercharger team before this “validator”? SMDH

Yep,"in any manner", they definitely should have reported it according to the law. Looks like there is a proposal to make traffic violations reportable. https://digitaldemocracy.calmatters.org/bills/ca_202320240ab3061 Frankly it seems just as likely the FSD is causing a bunch of unreported...

I like the panel alignment hypothesis. 1 cm accuracy explains a lot! :p Luminar says "This was offset by sensor sales to Tesla, which was our largest LiDAR customer in Q1 and comprised more than 10% of our revenue in the quarter." That doesn't preclude that much of the revenue was from...