d21mike
Active Member
“Once Tesla is able to lane keep .....” is no where near FSD. It has done that pretty well fir some time. NOA on highways and surface streets better than most drivers is required for FSD And we are still a long ways off.
-
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
I agree with you, but if you watch that video, it's getting a lot better. IMO, it's better than some old humans on those curves.
My point is that we can track FSD progress simply by the ability to lane keep. I believe once lane keeping is 99.999%, FSD will follow, and based on Tesla's progress so far, it's not that far off (1-2 years IMO), again, based on that video.
EVNow
Well-Known Member
Even prior to 24.4 I've been using AP for steep hills with curves. More difficult than what the video shows.
Anyway, saying FSD is done because it knows lane keeping needs a leap of faith.
ps : Basically what Tesla is trying to do is to solve "feature" to six 9s. When that happens, on that particular feature AP will be better than avg humans, but on other features it would still be terrible. There are hundreds of such features/scenarios to solve.
diplomat33
Average guy who loves autonomous vehicles
Uh no. There is a lot more to FSD than just lane keeping. There is responding to traffic lights and stop signs, changing lanes, making turns at intersections, yielding to other vehicles, avoiding pedestrians or cyclists and a hundred other tasks. The NN has to be able to handle all these tasks in addition to lane keeping to 99.99999% reliability. Plus there is driving policy too, rules like "if I see a police car on the side of the road, I need to move over to the passing lane". FSD is a lot more than just lane keeping.
[QUOTE="diplomat33, post: 3891178, member: 63496"rules like "if I see a police car on the side of the road, I need to move over to the passing lane". [/QUOTE]
Speaking as a human, I've never really learned how to do that when the other lane is 100% full.
Speaking as a human, I've never really learned how to do that when the other lane is 100% full.
EVNow
Well-Known Member
There’s a separate thread on that. Not much discussion though.
powertoold
Active Member
Y'all don't understand what I'm saying. I didn't say lane-keeping is all that's needed for FSD. what I was saying is that once lane keeping is 99.999% reliable, that means Tesla has developed a neural network that can adequately use vision to detect its surroundings and respond appropriately.
As of now, no, AP lane keeping is only 90-95%, still fails frequently, goes to opposing traffic, can't make sharp turns, etc.
Last edited:
diplomat33
Average guy who loves autonomous vehicles
That is not true either. The NN has to be trained separately for each task. So being good at one thing like lane keeping does not automatically mean that the NN will automatically be good at seeing say traffic lights or a duck crossing the street.
In his presentation, Karpathy mentions that they are training the NN for the following tasks:
- Moving objects (cars, trucks, motorcycles driving, pedestrians walking etc)
- Static objects (parked car, garbage can on the side of the road, traffic cones, pedestrian waiting to cross the street etc)
- Road signs (speed limit signs, stop signs, etc)
- Traffic Lights
- Lane lines
- Road markings (turn left arrows on road, pedestrian crossings etc)
- Environmental tags (residential area or highway, sunny day or raining)
Each of these have to be trained separately. Tesla has to get to 99.9999% independently for each of these tasks in order to do FSD.
I think it’s really exciting (no snark, I mean it) to see how good this can get on HW3. Looking forward to the next few month hoping this improves significantly.
willow_hiller
Well-Known Member
If my understanding of Tesla's NN stack is correct, some tasks are trained separately while some are naturally complementary and trained together. For e.g. reading signs and tracking lane-lines might fight each-other for loss if trained together, where static and moving object identifiers get better performance while trained as part of the same spine with different heads.
I'm sure this talk has been posted all over this thread by now, but Karpathy explained it really well in a talk a few months back: Andrej Karpathy | Multi-Task Learning in the Wilderness
diplomat33
Average guy who loves autonomous vehicles
Yes. Thank you for clarifying. The point still stands that reliable lane keeping does not automatically mean that the vision NN is good enough to do everything else. It's a delicate and complex process involving training a lot of different tasks and sub tasks as well. Right now, it seems like Karpathy's team is working hard managing that tricky process, to try to get the NN reliable in all the tasks necessary for FSD.
And maybe I am too optimistic (I have been accused of that many times) but I wonder if we could see a quickening of FSD features at some point when the NN reaches enough reliability in enough tasks?
powertoold
Active Member
Not true. AP as implemented right now is only for lane keeping, and right now, it is NOT more than 90-95% reliable as I said.
As of right now, there is no NN out there that is better than 90-95% at lane keeping. When I say lane keeping, I'm referring to moving forward within an appropriate part of the road, regardless of whether there are lane lines (as sober best humans can do with ~100% reliability).
What I'm saying is once Tesla AP can achieve lane keeping in 99.99% of situations, as a human can do, then FSD will follow.
We may never get there though, based on our current NN approaches.
diplomat33
Average guy who loves autonomous vehicles
I never claimed that AP was better than 95% at lane keeping. Not sure what I said that made you think that. All I said was that even if lane keeping was 99.999% reliable, Tesla would still need to get to 99.999% in all those other tasks like static objects, road markings etc before it could be FSD. And I was making the point that 99.999% reliability in lane keeping would not automatically guarantee the same 99.999% in those other tasks.
It sounds like you are rolling a lot of other things into "lane keeping". Are you including stopping at a red light or stop sign as part of "lane keeping"? Are you including lane changes as part of "lane keeping"? Are you including avoiding hitting an object in your path as part of "lane keeping"? It sounds like you are.
If you are, then I can see how you might see "lane keeping" as so close to FSD because you are essentially rolling in a bunch of other FSD features into "lane keeping". But generally, when people talk about "lane keeping", they are just talking about the ability to stay in the center of a lane, that's all.
willow_hiller
Well-Known Member
Been reading Asimov recently, it almost sounds like @powertoold is trying to define the first of Three Laws of Autonomy (as opposed to Robotics). But I'm not sure L5 can be simplified that way. That being said, an adaption of the three laws don't work so bad as fundamental self-driving rules:
- An autonomous vehicle may not injure a human being or, through inaction, allow a human being to come to harm.
- An autonomous vehicle must obey orders given it by human beings (rules of the road, emergency responder directives) except where such orders would conflict with the First Law.
- An autonomous vehicle must protect its own existence as long as such protection does not conflict with the First or Second Law.
diplomat33
Average guy who loves autonomous vehicles
Well, those 3 Laws could certainly be useful as philosophical guidelines to help write the driving policy of the autonomous vehicle. For example, the 1st Law would dictate that the autonomous vehicle should avoid hitting pedestrians, cyclists, etc... You still need to solve the technical part though. You still need to develop the software to make FSD work.
Although it does raise an interesting theoretical and philosophical question. Some folks have posited that true AI is the answer to L5 autonomy. In other words, they argue that we can achieve L5 autonomy when we develop a true AI that is capable of teaching itself how to drive. if such an AI were possible, could we simply give the AI these 3 Laws and basically tell it "you go teach yourself how to drive but you must obey these 3 laws when you do."?
powertoold
Active Member
My point is simple:
Until Tesla can achieve 100% reliable lane keeping in 99.99% of environments that humans drive, FSD is not possible.
As of now, they're at 90-95% with no guarantee they'll (or anyone for that matter) will reach 99.99%.
Once they do reach 99.99% lane keeping, I think they will have validated the technology and algorithms, and FSD will follow soon after.
So before we go gungho about discussing this and that, we have to realize that Tesla has yet to achieve highly reliable lane keeping.
Until Tesla can achieve 100% reliable lane keeping in 99.99% of environments that humans drive, FSD is not possible.
As of now, they're at 90-95% with no guarantee they'll (or anyone for that matter) will reach 99.99%.
Once they do reach 99.99% lane keeping, I think they will have validated the technology and algorithms, and FSD will follow soon after.
So before we go gungho about discussing this and that, we have to realize that Tesla has yet to achieve highly reliable lane keeping.
diplomat33
Average guy who loves autonomous vehicles
Well true. Lane Keeping is a prerequisite to FSD. So, obviously you have to have reliable Lake Keeping in order to have FSD.
Waymo has 99.99% reliable Lane Keeping already now. So it has already been done. I see no reason that Tesla can't get there as well.
Getting to extremely reliable Lane Keeping is no guarantee at all that FSD will follow.
powertoold
Active Member
Waymo does not have reliable lane keeping in 99.99% of environments that humans drive.
Hell, there's not a single video of a waymo car driving in anything other than flat city streets with well marked lanes.
diplomat33
Average guy who loves autonomous vehicles
Waymo is L4. I am pretty sure in the domain that they are designed for, they have 99.99% reliable Lake Keeping.
Plus, with the hardware that the cars have, I would be very surprised if they did not have 99.99% reliable Lane Keeping everywhere.
Similar threads
- Article
