Yep I know other companies have separate detections for open doors, but it’s either inaccurate or not yet available in Tesla’s code. Typically doesn’t show as a pedestrian on screen for me either.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Wiki MASTER THREAD: Actual FSD Beta downloads and experiences
- Thread starter jebinc
- Start date
old pilot
Member
Yesterday when I made a drive along a route I use to see how new iterations of the FSD perform, I came to a T intersection where I am in the right lane at the top of the T and am making a right turn into the vertical part of the T, in past drives there have been no cars stopped (at a stop sign) at the top of the T in the right lane, and the car moved slightly left to make the right turn and always crossed the yellow center line when completing the turn. Yesterday, for the first time, there was a truck stopped at the intersection in the right lane and the car could not exceed or cross over the yellow line without hitting the truck so the car went several feet to the left, crossing the yellow line (fortunately there was no oncoming traffic because it was a very abrupt move), before attempting the turn, I assume it did this to insure it would not cross the yellow line when completing the turn at the top of the T. I disconnected put the car back in its lane at the top of the T and made the turn, while staying in my lane at both the horizontal and vertical portions of the T intersection. I am becoming convinced that the person, who is doing the programming, is either a poor programmer but a good driver or a good programmer and a poor driver (I think the latter is more likely). It is clear, at least to this driver, the car is aware of its circumstances and is moving left to make a right turn not because if is necessary (its not an 18 wheeler) but because it is programed to do so for no good reason.
EVNow
Well-Known Member
In any case - the car goes too close to the parked cars. If someone opens the door, it will hit the car. They need to give more space, if there isn't anything on the left.
EVNow
Well-Known Member
That is not how it works. The car chooses a path based on cost-optimization based on hundreds of simulations.
What Is Monte Carlo Simulation? | IBM
Monte Carlo Simulation is a type of computational algorithm that uses repeated random sampling to obtain the likelihood of a range of results of occurring.
www.ibm.com
The question is - what parameters and weightages / costs have they used. Seems to me it is making wider turns in 10.5 to avoid curbing (a common complaint before) i.e. they cost associated with curbing is much higher than cost associated with cross the yellow line.
ps : Having specifically watched for this - I should say almost everyone crosses the yellow line when turning if there are no vehicles around - esp. in the smaller local streets.
They do, but its really bad behavior.
Human drivers are so used to crossing the yellow line on a left turn that they often cut off people doing the other way. Or they have to correct mid turn because they entered too shallow, and they have to correct,.
For human drivers I think its more of a speed thing. Like I'll do the same thing if I'm in a hurry, and if there isn't anyone there.
But, I expected an automated driving system to always have a perfect turn.
EVNow
Well-Known Member
Absolutely - one of my expectations (say a year or so back) was that FSD would make perfect turns. Afterall it can calculate what the best curve to get to the middle of the target lane was.
The way they have setup the Monte Carlo simulation, it is not yielding the best path for this one "edge" case.
old pilot
Member
I do not have expertise, that you and others hear have in this area, so I am sure you are correct in what you say, however, what I am saying is that it is possible to make these turns safely and without damage without crossing the yellow line either on the entry to or the exit from the turn. I would agree that most drivers will exceed or cross the yellow centerline at times when completing a turn and there is no oncoming traffic but that does not mean you must swerve or move left to make a right turn. It is a very poor driving habit and specifically addressed in many driving manuals.
They don't quite use that...They use a classical discrete search algorithm with continuous optimization function applied on top..That is not how it works. The car chooses a path based on cost-optimization based on hundreds of simulations.
What Is Monte Carlo Simulation? | IBM
Monte Carlo Simulation is a type of computational algorithm that uses repeated random sampling to obtain the likelihood of a range of results of occurring.
The question is - what parameters and weightages / costs have they used. Seems to me it is making wider turns in 10.5 to avoid curbing (a common complaint before) i.e. they cost associated with curbing is much higher than cost associated with cross the yellow line.
ps : Having specifically watched for this - I should say almost everyone crosses the yellow line when turning if there are no vehicles around - esp. in the smaller local streets.
You are getting that mixed up with what they plan (wish) to do aslong side that which is to add a NN planner that uses Monte Carlo Tree Search (MCTS) and a NN that helps efficiently explore the tree.
EVNow
Well-Known Member
Absolutely. Just explaining how it works and possible cause. Really has nothing to do with how good the programmer is in coding or driving. Its also probably a team and not a single individual.
Good thing is - they should be able to fine tune to get the optimal results in all conditions.
EVNow
Well-Known Member
I've to watch again - but it sounded like they were using Monte Carlo so to reduce the number of simulations / searches they had to do. May be that was WIP.
Either way - the core idea doesn't change. They are not "hard coding" the path for every situation.
ps : After a bit of reading - looks like what they were/are using is a hybrid approach. MuZero based NN for planning is probably in development.
Deep Understanding Tesla FSD Part 3: Planning & Control
From Theory to Reality Analyze the Evolution of Full Self-Driving from Tesla AI Day
saneryee-studio.medium.com
Last edited:
But not too much space that it rolls down the middle of the street as it used to. I agree it feels like it can get too close after whatever they changed in 10.5. But it fully rolled down the wrong side of a wide 2-way unmarked road earlier this week anyway, with > 50% of the car on the wrong side of what a human driver would consider to be the virtual dividing line. No pedestrians or any other moving vehicles on that block. Consistency is something they really need to nail down haha
EDIT sort of goes towards the discussion above, the only cause I could think of for driving on the wrong side was an upcoming right turn
dandrewk
Active Member
One of the goals in generating the path is comfort which they say is smoothness. So they are probably trying to minimize curvature. By moving through the outermost part of both lanes, the circle that joins them is as large as possible, and it's curvature is minimum. It's not necessary to make the turn, but it's part of their overall goal.
Last edited:
old pilot
Member
I understand the goal may be smoothness but it should be safety first and moving out of your lane into oncoming traffic is not the answer. It is possible to accomplish a right turn while staying in your lane and still be smooth and comfortable. It requires moving forward at the proper speed to allow for the delaying of the turn until the left front wheel and the right rear wheel are in the proper position to allow for a sharper turning radius. Others here have stated that it is being done (moving left) to stop the curbing of the rear tire but once again this can be accomplished by using proper technique.
Delaying the start of the turn requires a sharper turn which would be less smooth.
I don't think they are adjusting the weights for smoothness when they are turning, and we know double yellow lines are more of a guideline for the path planner. So the combination of low weight on the double yellow rule combined with a desire for smoothness lets the car intrude on other lanes. If there is another car in that lane, the weight to avoid collisions should cause the car to stay in its own lane.
It looks like the NN generates the heuristic that provides discrete values for the mcts to minimize. I don't think the NN affects the operation of the search algorithm.
For intersections with obstructed views to the left FSD needs the car to be perpendicular to the intersection so the front cameras can see on coming traffic. (B-pillar is obstructed). When making the right hand turn and needing to avoid crossing the center line FSD has no choice but to be jerky. I have this situation every day and I don't see it getting resolved with the current camera setup.
old pilot
Member
The G forces are controlled by the speed of the turn, so slowing the speed will keep the turn smooth and comfortable and still keep the car in its lane. I think I understand your position that this has to do with how the computer is programed to meet certain goals. I don't care about that as the passenger in the car, I want to feel safe and because I cannot interact with the computer as I would with a human driver the computer driver must be perfect in it execution of maneuvers. Your last sentence, I think, is what causes me the most problem with trusting the computers driving, that being the unknown, the use of the word "Should" If you were able to say with conviction "Will" I would feel safe. I hope this explains my position, its about trust and knowing not faith and hope.
There is also a goal of completing paths in a minimum amount of time. So the resulting path is a compromise among competing goals, and they can't all be perfectly accomplished simultaneously.
I think all these issues will ultimately be solved, but they shouldn't spend effort optimizing every stage along the way. It's better to get a complete solution before going back and optimizing. Some of the current algorithms may not survive to reach the final solution.
I originally wrote "would," but we know that it can do the wrong thing at the worst time, so don't get comfortable.
Last edited:
I would be curious as how you think Tesla will solve the first item above when you have to make a hard 90 degree turn without crossing the line on a single lane road? If FSD would approach the intersection pointed slightly to the right like all human drivers I'd see no problem but with an obstructed B-pillar camera it seems like staying perpendicular and having to creep into the crossing road is mandatory.
Similar threads
