Would be nice to have a debugger's video review option. So you can find out, and learn what your car is doing.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
EVNow
Well-Known Member
If its a perception problem - then they may need more training or better optimization.
My guess is - its a planning issue. In YouTube videos I've seen it drive in between yellow lines, but usually near the turn when the lanes are backed up (which is what people do as well).
AlanSubie4Life
Efficiency Obsessed Member
It's actually always illegal to drive in that area. It's treated as a solid barrier that cannot be crossed. So I was breaking the law when I allowed the vehicle to drive in that area this morning. If it's not coded that way in FSD, it seems like a major oversight (it's not a complicated rule, and this is California so it's a very common case).
Lane Control - California DMV
Examples of lane markings (1) Solid yellow line: No passing if solid yellow line is on your side.(2) Double solid lines: DO NOT pass.(3) Broken yellow line: May pass if movement can be made safely. Line Colors Solid yellow lines mark the center of a road used for two-way traffic. Broken yellow...
www.dmv.ca.gov
EVNow
Well-Known Member
Ofcourse, its illegal to drive there. Just as it is illegal to drive even 1 mph above the speed limit
Another NHTSA investigation idea:
"Hey Tesla, your cars are programmed to allow California rolling through stop signs.
Please provide the technical and legal basis for this maneuver, and list any other violations of the Motor Vehicle laws that you have allowed in FSD Beta"
AlanSubie4Life
Efficiency Obsessed Member
Ofcourse, its illegal to drive there. Just as it is illegal to drive even 1 mph above the speed limit
You may be thinking of a case like this, where it is completely legal to enter the area between the yellow lines - for example, when traffic is backing up? It's totally different. (In this specific case below, you'll find that if you attempt to use the area to get into the turn lane, you'll end up with an unpleasant surprise - and the center turn lane reverts to solid double yellow before that point - and technically you can only drive there for 200 feet).
You will eventually be ticketed for driving illegally if you choose to enter the much different area with solid double-yellow demarcation. It's really not ok to do so. It's marked that way for a reason - it can abruptly end with boulders, shrubs, barriers, curbs etc.
And by the way, on the prior road (Spring Canyon) that I pictured, it's actually not illegal (in the enforceable sense) to drive above the speed limit - because there is no up-to-date speed study that has been conducted. They cannot enforce the speed limit on that road under California law (part of the "speed trap" law). The neighborhood/community chooses to keep the limit posted at 45mph, and NOT have a new speed study done (rendering it unenforceable), because it would result in the speed limit being raised to 50 or 55mph, which may be less safe. It's better to just have people driving at 55 or 60mph (55mph seems fairly typical) rather than 65 or 70mph. Slippery slope. End digression...
Last edited:
I'm getting really inconsistent behavior re: stop signs where I'm at.
Sometimes it'll do a complete and total stop, other times it just slows down to a crawl and rolls through them. The car does "see" the stop signs that it rolls through.
Smooth braking when approaching traffic lights is another head scratcher, too. Sometimes it's completely smooth, other times it's bad enough to make my wife car sick.
One thing it has shown consistency with, so far at least... mad lad acceleration out of roundabouts.
stopcrazypp
Well-Known Member
Well, putting aside this is not the federal government's jurisdiction (probably CA DMV cares more) they can probably ask the same question for every L1/L2 system that allows you to set a speed above the posted speed limit (or max speed limit), as CA's law is even when you are 1 mph above the given speed limit, it's technically illegal.
powertoold
Active Member
Many of the issues we’re experiencing with 10.2 are related to perception: jerky turns, fuzzy road geometry (fluctuates pathing), curb scares, crossing lines mid-turn, etc. Given perfect perception, I bet the planner would do a better job.
Tesla has been improving the perception, and we know they have larger NN models that have better perception, but will they be able to fit the necessary perception models into hw3? We don’t know if this is a labeling limitation or an optimization limitation. It seems to be a labeling limitation based on what Ashok said during AI Day.
Tesla has been improving the perception, and we know they have larger NN models that have better perception, but will they be able to fit the necessary perception models into hw3? We don’t know if this is a labeling limitation or an optimization limitation. It seems to be a labeling limitation based on what Ashok said during AI Day.
Yes maybe they can look at how other systems can be misused to exceed the law - to the extent "dumb" systems can be regulated. For better or for worse we will get more nanny-state rules. That's what happens when someone "ruins it for everyone".
Or Tesla could just remove California stops and pre-empt this problem. We know it's a purposefully engaged parameter from those hacker reports. It's so simple for Tesla to deprogram it...
AlanSubie4Life
Efficiency Obsessed Member
I know just the solution for that! A fundamental rewrite.
Seriously, I agree that it seems that perception still has a lot of issues. The question is, will they have to entirely re-arrange their perception engine at any point? Or are they at a point where they can just add features with the existing architecture and have it not be constrained by hardware at any point in the near future?
I'm not even sure what benefit it provides since FSD Beta does such an awful job at stop signs, anyway. It'll jerk on the brakes early, roll freely for a bit, then jam them on again, possibly stopping prematurely way before the line, or possibly just rolling through the stop sign.
It seems like it wouldn't be asking too much to come to an aggressive & bold, yet predictable, smooth gradual full stop at the line, using only regen, then continue to creep from there as needed (often there is need to creep). How hard can it possibly be to stop reliably at a certain point on the road and minimize jerk in doing so? (Apparently pretty difficult?)
Knowing the vehicle characteristics and accelerator/regen state, and knowing the behavior of velocity with that input, they should be able to figure out the slope, etc., and get a pretty good idea of how to optimize the stop. Humans are SO good at this and we don't even need to use math. Oh well.
Last edited:
Yeah I'm sure it drives great inside their Unreal Engine simulation. I don't know but the planner seems pretty naive to me, from AI Day they didn't spend much time on control/planning, but it sounds like the path planning is just a tree search that tries to minimize some cost function, like those chess playing AI programs. This is not how humans drive.
Here's an example of wonky path planning, the car is supposed to keep traveling straight for the next few miles at least. But the planner wants to turn left into a driveway for just a single frame, like 10 ms. There's nothing obviously wrong with the BEVnet predictions, it shows the lane continues straight ahead and there is a driveway. And so the steering jerks for a brief moment and then corrects itself. For a brief moment the cost of traveling straight exceeded the cost of swerving to the left at 30mph, which just seems totally broken to me.
AlanSubie4Life
Efficiency Obsessed Member
When the tree search cost function and the BEVnet disagree, which one do you believe???
I agree that it seems they have some work to do here. It seems like they may not have adequately merged the trip planner and the cost function somehow. Or they have their weights wrong, or something. (To be clear, again, I have no clue.) The cost function isn't even working very well when it comes to its fundamental purpose of driving smoothly, either. Hopefully they get some stability in their development soon and can go back and focus on optimizing what they have. It seems to me they're in vicious cycle where they keep getting pushed to add features and label objects and respond to them, but the existing features still need a lot of work, but there's no one to go back and work on them. Maybe they can go back and work on fundamentals for a couple months, now that they have a very limited beta release of FSD Beta?
What would help is if they forced all the engineers (and perhaps more importantly, the engineering managers) working on FSD Beta to drive a certain number of miles using FSD Beta each day.
powertoold
Active Member
I don't know how long you've been following fsd / the videos, but what you're seeing with the wonky path planner is a result of perception uncertainty and optimizing for safety.
We saw this with V9 and Chuck's unprotected lefts, where the route was to turn left, but the planner decided to turn right almost every time. This is because the planner is limited by the certainty of the perception. If the perception tells the planner for 10ms that you're in a left turn lane, the planner will attempt to turn left. That's why we see the planner path jump all over the place at times.
This sort of planner (that optimizes for certainty and safety) has pros and cons. For example, we saw in AI DRVR's SF video in V10 that the planner is able to adjust to unsuccessful lane changes:
Last edited:
EVNow
Well-Known Member
Unless - the visualization shows some kind of smoothing applied to the predictions (because of all the gripes about dancing cars) but the cost optimizer uses real-time non-smoothed predictions ?
Cost optimization isn't easy either - and may sometime give bad results (like super aggressive right turns, when the road is empty). But it can get better over time - but then it becomes more & more difficult to make any change (just like any procedural code).
I also wonder whether FCW is being caused by prediction from NN or planner or some other procedure code they haven't adequately updated for Vision.
ps : Talking about adding new features without fixing existing bugs ... "single stack" is probably the worst offender.
Last edited:
This is a good argument for having a more end-to-end model with these sorts of tasks. If I had to put myself in the shoes of the Tesla path planner, and instead of driving based on what's out the windshield, I'm driving on the BEVnet/lanes network output, I'd probably do a pretty crappy job. I'd be scared to do anything and would be very skittish, all my rich visual cues are gone, and I'm half-blind driving through this stark vector space of road/not-road. I think I could do a better job than the current planner, but not by much.
powertoold
Active Member
2 cars doing exactly this, lol:
Gali let's 10.2 hit a curb so he doesn't have to disengage...... "looks like we're batting 1000 so far" (after he hit the curb, lol!):
Last edited:
AlanSubie4Life
Efficiency Obsessed Member
Yeah, given its current limitations (specifically can't reliably identify curbs apparently) I wouldn't have FSD doing what those cars are doing quite yet. As mentioned, double (actually quadruple yellow specifically) yellow can presage curbs, boulders, shrubbery, trees, etc. Yet, apparently it still ignores even double yellow (which is sometimes legal to cross of course).
"I've noticed it's a lot less jerky than previous versions....very human-like with the way it slows down and speeds up."
Dickens' “It was the best of times, it was the worst of times” really seems to apply to 10.2 from what I've experienced, by what people are posting here, and by what is being seen in many of the FSD beta videos.
This video from Ryan Shaw seems to capture that wide range from the surprisingly well executed moves to the white-knuckle crap-your-pants moments. While the video shows that whole range it's really his ending statement that seems to reasonably reflect where we are as of today. The optimist in me looks forward to what 10.3 might bring.
This video from Ryan Shaw seems to capture that wide range from the surprisingly well executed moves to the white-knuckle crap-your-pants moments. While the video shows that whole range it's really his ending statement that seems to reasonably reflect where we are as of today. The optimist in me looks forward to what 10.3 might bring.
Last edited:
Similar threads
