Just got to say — Tesla really needs to work on its allowance of merging traffic when your in AP on the highway. It is a complete AH to people trying to merge on. Almost hit the merging car because AP would keep creeping forward instead of giving the car to the front right of me enough room. Their rear tires were at my front right quarter panel probably 6 inches away before I hit the brakes. Almost seems like a blind spot for AP
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
AutoPilot and merging traffic
- Thread starter QS01
- Start date
jdcollins5
Member
Why do you think AP would know that you want the car in the right lane to merge in front of you? When in AP it is following the car in front of you at a set distance. If you want to let a car merge, lift the gear stalk to disengage and then re-engage after the merger.
it seems like you need to understand how AP works.
This is going to always be the problem with AP and all auto driving system. They will always suffer from being subjectively too aggressive or not aggressive enough. It's a neural net variation of the old: "Everyone going slower than me is a dumb a$$ and everyone going faster is an a$$ hole".
PianoAl
Active Member
I have experienced the same thing as the OP. Perhaps three times I've had to take over.
I know it's not FSD, but I was surprised that my M3 let a car get that close without slowing down or changing lanes to avoid it. It was as if it didn't see it at all.
If someone ahead of me in a lane slows down, my M3 will slow. Shouldn't it do something similar for someone merging in front of me?
What happens if you're in the left lane and a car attempts to merge into your lane but too close to you?
I know it's not FSD, but I was surprised that my M3 let a car get that close without slowing down or changing lanes to avoid it. It was as if it didn't see it at all.
If someone ahead of me in a lane slows down, my M3 will slow. Shouldn't it do something similar for someone merging in front of me?
What happens if you're in the left lane and a car attempts to merge into your lane but too close to you?
I agree -- I would fully expect the M3 to slow down when there is a vehicle that close and not keep forcing the issue. I am fully aware it's AP and not FSD, however AP stays in the lane and my expectation is that it should handle merging traffic while staying in that lane. If this isn't something AP will do, it should expressly be called out IMO by Tesla (like the lane change with turn signal is) as a feature only available to FSD.
PianoAl
Active Member
And the worst part was that before I had to take control, I said to my wife, "Okay, now watch how it handles this car merging in."
jdcollins5
Member
Once again, you need to understand how the car is designed to work and not how you think it should work. For your sake and those around you.
Unless you “know” how it works, don’t “assume” how it works.
Unless you “know” how it works, don’t “assume” how it works.
and again... I realize it doesn’t handle merging for crap so yeah I monitor that. That doesn’t change my original point that Tesla needs to work out the kinks with merging traffic. That’s all my main point was.
That has happened occasionally to me too, in very stupid/annoying/unsafe fashion. Now I make sure I disengage AP when in the right lane and driving past an on-ramp and there’s another car anywhere nearby.
This is one of those instances where the Tesla engineers try to make Autopilot “smarter” in response to complaints like the OP’s, but in reality they just make it worse.
AP in general is kind of bad at detecting merging lanes today; sometimes you can watch the car you're about to merge into get correctly recognized and then ignored a few times in a row. Pretty often it detects everything correctly but once the merge is close enough it's like the car forgets that there aren't 2 lanes anymore.
Exits/entrances seem to be pretty well-mapped around here so it isn't as confused about those.
Exits/entrances seem to be pretty well-mapped around here so it isn't as confused about those.
Why is there always an implicit assumption that AP is “done”? It’s constantly improving but doesn’t handle lots of scenarios ... yet.
I actually find it’s merge behavior interesting. It appears from my observations to handle merging traffic well if that traffic has its signal on. No signal and the car doesn’t handle it much at all.
I actually find it’s merge behavior interesting. It appears from my observations to handle merging traffic well if that traffic has its signal on. No signal and the car doesn’t handle it much at all.
My experience is that I've had the car slow once to allow a car to merge from an on-ramp. Locally, most of our on/off traffic uses a temporary right lane that is lined off. In that case the car totally ignores any attempt to change into it's lane, even if the other car is signalling and moving towards my lane. Our cars certainly seem to have a blind spot for the front quarters.
WarEagleX
Member
Even fully FSD capable cars will have issues with merging until the advent of car-to-car communications. Merging is perhaps one of the most complex dynamics of driving because there is a lot of "instinct" involved on your part as you act on subtle hints of the oncoming traffic to anticipate the intentions of the oncoming drivers - and we do this extremely well where machines don't.
Think about it. When entering a highway, when there is medium to heavy traffic, we look to our left and in our side mirrors briefly, and immediately assess a flood of information to note what the traffic is doing and decide what we should do based on that brief look. We note if there is a car in the right lane for which we should yield, or a car approaching quickly that we anticipate will be a danger to us by the time we get to our intended merge point. We may slow down at different levels accordingly, or we may accelerate. Then we also realize and anticipate that a lot of people won't allow us in, and we can tell by how they're driving, so instead of slowing down, they speed up to prevent you from merging ahead of them. Some cars will move to the left to let you in. Some cars will flash their lights to let you know that they are letting you in so we use that as a cue to speed up and merge aggressively so as not to inconvenience them - and a wave to say thank you. Some cars don't take any actions at all, as if you're not even there, and it is up to you to figure out how you should act because they don't speed up, slow down, or move over. There may be a truck coming that we want to get ahead of, so we accelerate or it's coming too fast so we decelerate because trying to merge ahead of it may exceed our danger tolerance, even though we probably could have made it. We process all of these subtle variables in milliseconds, come up with a plan, and then execute.
The problem with AI and ML is that it cannot process human "intent". It also cannot anticipate random variables simply by subtle hints of the vehicles and drivers around it like we can. This flood of information that humans process though the millions of years of evolution of our species are impossible for the current state of AI to recognize much less understand.
This is where upcoming car-to-car communications systems from autonomous cars come into play. When merging, your car will broadcast to adjacent cars of its intentions and those cars will either speed up or slow down and communicate their intent back to your car at which time it will take an appropriate action based on explicit and predictable information. Your car no longer has to read human intent and it is now 100 certain on the action it should take.
So, until that time, all of us Tesla drivers (mine won't be delivered until Feb 11th) would do well to understand the current level of AI and ML. If you don't, then two things will happen. First, you will be constantly disappointed that your car can't do the things which you take for granted as a human. And second, you will become a statistic with the top of your car - and most likely your head - missing after it drives itself under a tractor-trailer at 65 miles per hour.
Think about it. When entering a highway, when there is medium to heavy traffic, we look to our left and in our side mirrors briefly, and immediately assess a flood of information to note what the traffic is doing and decide what we should do based on that brief look. We note if there is a car in the right lane for which we should yield, or a car approaching quickly that we anticipate will be a danger to us by the time we get to our intended merge point. We may slow down at different levels accordingly, or we may accelerate. Then we also realize and anticipate that a lot of people won't allow us in, and we can tell by how they're driving, so instead of slowing down, they speed up to prevent you from merging ahead of them. Some cars will move to the left to let you in. Some cars will flash their lights to let you know that they are letting you in so we use that as a cue to speed up and merge aggressively so as not to inconvenience them - and a wave to say thank you. Some cars don't take any actions at all, as if you're not even there, and it is up to you to figure out how you should act because they don't speed up, slow down, or move over. There may be a truck coming that we want to get ahead of, so we accelerate or it's coming too fast so we decelerate because trying to merge ahead of it may exceed our danger tolerance, even though we probably could have made it. We process all of these subtle variables in milliseconds, come up with a plan, and then execute.
The problem with AI and ML is that it cannot process human "intent". It also cannot anticipate random variables simply by subtle hints of the vehicles and drivers around it like we can. This flood of information that humans process though the millions of years of evolution of our species are impossible for the current state of AI to recognize much less understand.
This is where upcoming car-to-car communications systems from autonomous cars come into play. When merging, your car will broadcast to adjacent cars of its intentions and those cars will either speed up or slow down and communicate their intent back to your car at which time it will take an appropriate action based on explicit and predictable information. Your car no longer has to read human intent and it is now 100 certain on the action it should take.
So, until that time, all of us Tesla drivers (mine won't be delivered until Feb 11th) would do well to understand the current level of AI and ML. If you don't, then two things will happen. First, you will be constantly disappointed that your car can't do the things which you take for granted as a human. And second, you will become a statistic with the top of your car - and most likely your head - missing after it drives itself under a tractor-trailer at 65 miles per hour.
Similar threads
