powertoold
Active Member
Lidar doesn't help much with lane-keeping. It's mostly vision based, and again, our vision tech right now can't do 99.99% lane keeping. Doesn't matter how many sensors you have.
-
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
Lidar doesn't help much with lane-keeping. It's mostly vision based, and again, our vision tech right now can't do 99.99% lane keeping. Doesn't matter how many sensors you have.
diplomat33
Average guy who loves autonomous vehicles
I know. But Waymo cars have plenty of cameras too and Google has been working on vision longer than Tesla. Also, Waymo cars use HD maps. If the vision is not good enough to do Lane Keeping, the car fall back on the HD maps to do Lane Keeping.
I would be shocked if Waymo cars did not have 99.99% reliable Lane Keeping at this point especially since according to the official DMV Report, Waymo only has 1 disengagement every 11,000 miles and none of the disengagements are caused by poor Lane Keeping.
powertoold
Active Member
No. I will say with certainty that waymo does not have 99.99% lane keeping. If you agree that lane keeping is reliant on vision, then it follows that vision is reliant on data, and waymo doesn't have the data.
EVNow
Well-Known Member
How do you define 99.99% reliability? I'm fairly sure Waymo does better than 1 error in 10k driven miles in their geofenced areas. Probably Tesla too.
BTW that is better than humans. We cross over the lane much more frequently.
powertoold
Active Member
Maybe we're talking about 2 different things.
When I say FSD, I'm mostly referring to level 5, best human level routine driving.
And if we're talking about best human level driving (i.e. sober human with all faculties intact), no, waymo is not anywhere close, neither is Tesla, and I'm just referring to lane keeping. We humans can lane keep in pretty much all environments if asked to do so, like say in a test or something.
Sure, waymo can geofence a few blocks and drive it in circles. If you're referring to that, yes, it's > 99.99%.
Given that we only use about 140 million neurons per hemisphere for vision it’s truly remarkable. Even watching the Tesla Paris FSD video on my iPhone screen allows me to identify the drivable surface with effectively mono vision using a tiny area of my field of view.
EVNow
Well-Known Member
You are still not saying what your definition of that four 9s is.
Humans can keep to within lane lines in a test of few miles. But over a year of driving 10k miles, they don't, even with proper markings.
This is why definition of that 99.99% is important to establish what we are even talking about. Is it errors by miles driven, errors by type of roads, errors by types of markings ?
My P3 can’t keep it in the lane for about 30% of my commute. There’s a reason why I’m so skeptical. And no, it’s not just my P3. The test drive vehicle had similar performance.
Looking forward to experiencing HW3 NN lane keep performance.
diplomat33
Average guy who loves autonomous vehicles
What roads do you drive on? Lane Keeping on my Model 3 is excellent on the roads I drive on.
IH635 and a few miles on a 6 lane divided city road.
I’m starting to wonder if the HW3 port of the HW2.5 NN is the problem?!
diplomat33
Average guy who loves autonomous vehicles
Probably. I think Elon even said that the HW3 port would not be optimized at first and would be a bit worse than AP2 until it got optimized. When we get the true AP3 NN, it will be much better.
powertoold
Active Member
If you take the best human drivers and ask them simply to stay within the lane in any example driveable road in the world, they'll be able to do it for at least a few minutes, with curves and crossing intersections and all. No autonomous car can do that right now, not even close. Pretty simple to understand IMO.
Just from that alone, we can see how far FSD is.
EVNow
Well-Known Member
You keep moving the goal post and not giving a workable definition that actually let's us calculate the reliability.
powertoold
Active Member
If I were to define the exact parameters, I'd be writing a whole paper.
Since this is a board and not science journal, I like to make it simple: Tesla AP can't even lane keep reliably right now. Until it can, fsd is still far away.
diplomat33
Average guy who loves autonomous vehicles
Not true. Waymo cars can do that now. Waymo cars drive on various roads, including handling intersections and only have 1 disengagement per 11,000 miles. So they can clearly do it for more than a few minutes. Autonomous tech has progressed far more than you realize.
diplomat33
Average guy who loves autonomous vehicles
I guess it depends what you mean by reliable. But I would rate that statement as false.
EVNow
Well-Known Member
So, is your car HW3 ?
I drive on complicated roads with lots of curves in the suburbs of Seattle. Car keeps to the lane except when
- the lines are badly faded
- curving road with merging, splitting lanes
- intersections where the leading car turns
Even in these cases most of the time the car does ok.
Are you saying in simple well marked roads your car doesn't keep lanes 30% of the time ?
EVNow
Well-Known Member
No, you don't have to write a paper. It is a simple question - how do you calculate that 99.99%.
If you are purely talking qualitatively, say so.
Yes, it’s HW3.
The lane markings aren’t super clean when it fails. It works well when the lanes are well marked but lane detection deteriorates significantly when the markings are a bit worn.
There’s one stretch where it consistently fails during autosteer. It’s the right lane of a six lane divided stretch with dashed lane markings that are a bit worn. It always starts to lose its lane confidence without disengaging autosteer and drifts from the right lane into the middle lane.
EVNow
Well-Known Member
Similar threads
- Article
