EVNow
Well-Known Member
Its not clear v10 has been created primarily for HW3. That might be post v10 after Tesla starts upgrading old cars.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Autonomous Car Progress
- Thread starter stopcrazypp
- Start date
-
- Tags
- Autonomous Vehicles
powertoold
Active Member
We can be quite certain that V10 will eventually be feature complete for FSD. It's unlikely that Tesla will finish or release version 11 by the middle of 2020.
That being said, we have to expect that V10 will be a big step up for AP.
Can't they release those fixes / improvements now? Ghost braking like crazy here...
diplomat33
Average guy who loves autonomous vehicles
I'm not sure it quite works that way. It's not like there is a line of code that says "phantom_braking: enabled" that an engineer can just go in and disable. Tesla uses a NN that controls what the car sees. Tesla is continually training the NN to try to get it more and more reliable. Now, hopefully, the training pushes the NN in the right direction where it sees things better which results in less phantom braking.
One area that I hope V10 will fix is the delayed braking from cross traffic. Right now, when AP sees a car cross in front of you, it will brake. But since the car is moving sideways, the braking happens after the car has already moved out of the way. Hopefully, when Tesla implements cross traffic recognition as part of "automatic city driving", this problem will be fixed.
I know that, I am a software engineer myself, who has worked on computer vision. But the improvements should be incremental, so why not release it along the new versions instead of waiting for V10? I am not sure they get better, they surely didn't from V8 to V9.
The NN is a lot of different nets interconnected together. You can't easily improve one section without impacting others. They have to pick a balance/ local maximum of overall performance. Then to make the release they run that through their internal test fleet, external early acccess group, and then roll out.
Tesla has a history of major releases in August along with smaller tweaks during the year.
They have several tasks. (Andrej Karpathy | Multi-Task Learning in the Wilderness ) And a lot is not NN but "classical" programming where they do the decisions, filtering, the actual fusion etc. So I see no problem there. Did you notice a HUGE jump in improvements from V8 to V9? The "tweaks" can actually make a huge difference, but it has gotten worse.
Phantom braking will get a lot worse before it gets better
I am also aware that it could get worse before it gets better, but they should then add a thicker layer of "classical" code on top to massage and filter the data and do decisions and then cut down on that layer when it's due. I know they are not crappy programmers, but they may have a "NN first" philosophy with as a thin "Software 1.0" layer as possible. I'd say they should have a thicker Software 1.0 when the NN is confident enough. This is evident in the "dancing cars" that they have a pretty direct representation of what's actually going on. On one side it's honest, but on another side it's not a good user experience.
Last edited:
EVNow
Well-Known Member
Did you mean when the NN is not confident enough ?
I think it depends on how much resource the "thickening of software 1.0 code" would take. Since it is throw away code - and if it takes significant resources it is not worth it. Better to put those resources in improving NN. Afterall a short term hack is still just that - a hack.
Yes
Thing is - let's face it - that the "throwaway code" has to be there for a really long time. So in order to have a good user experience they should put more effort into it. I know they want to get there as fast as possible but the experience in the meantime is pretty bad.
They should hire a few people focusing on that layer.
Another problem are the loss functions. I guess that's why we see regressions all the time.
Let's hope that V10 brings a NN that blows our socks off.
EVNow
Well-Known Member
I personally agree - but Tesla apparently thinks otherwise. I can see Musk shooting down such ideas that are based on pessimistic reading of the situation.
powertoold
Active Member
Will FSD ever be possible with edge cases like this? It doesn't even seem like an uncommon case:
Daniel in SD
(supervised)
Haha. That is not an edge case nor is it anywhere close to the most difficult driving situations. You should watch some of the Cruise automation videos YouTube.
Ha,
Tesla has a history of major expected released in August that actually land in Sep/Oct, and sometimes later.
diplomat33
Average guy who loves autonomous vehicles
That's not an edge case. And yes, I am sure Autopilot will be able to handle this at some point.
AP is not FSD yet and is still a work on progress so you will be able to find a bunch of cases where AP still fails. Finding an example where AP fails does not mean that FSD is not possible, it only proves that AP still has work to be done.
Thanks for the correction. I'd picked up August on some previous search, but reviewing the blog posts I see I was wrong.
It's fine that it's not FSD yet, but it wouldn't hurt to have a couple of lines of code that acted on certain scenarios. I feel that the "software 1.0" code is minimal.
Yep. @powertoold, see Cruise like the double-parked, 1,400 left turns and SF maneuvers videos. Also look at Why testing self-driving cars in SF is challenging but necessary, or just the 4 examples if pressed for time.
Furthermore, this particular case is almost a non-issue for Cruise or anybody else that uses HD maps. They don't need to see the lane lines or even the pavement with the cameras to know where the lanes are. The maps tell them exactly where the lanes are, so long as they are able to locate themselves in the map correctly. This is only a challenge for Tesla and other non-map-based lane keeping systems (none of which other than Tesla make any claims about being fully autonomous ever).
How did AP1 in 2016 perform in situations like that?
It’s 2019. What is going on??? Why is this still not working better?
diplomat33
Average guy who loves autonomous vehicles
I don't really like these dashcam only videos that don't show what was on the driver display. Heck, we have no proof that AP was even on. And if AP was on, were there warnings on the screen? Did AP lose the lane lines? Did AP shut off? We are lacking a lot of information that could help us diagnose maybe what happened.
Similar threads
- Article
