roblab
Active Member
My car changes the speed limit JUST AS I pass the signs. I don't remember that it just changes for no visible reason, like a GPS notification.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Phantom braking is the biggest issue with AutoPilot.
- Thread starter WilliamG
- Start date
My car changes the speed limit JUST AS I pass the signs. I don't remember that it just changes for no visible reason, like a GPS notification.
Knightshade
Well-Known Member
Which car? (you have a '15 S and an '18 3 in your sig)
AP1 cars read speed limit signs.
AP2+ cars do not, they use maps/speed limit database/GPS location
thodgdon66
Member
I’ve got two 2019 M3P’s, both HW3, mine has 8600 miles, wifey about 2200
Just on the last update (both cars on 2020.4.1), mine (black) has experienced a pretty aggressive phantom braking symptom twice. I did not pay attention to the situation specifics other than it was highway and in a slight curve left.
I have seen zero phantom braking since whatever revision we were on in November, and never previously was it this aggressive. Keeping an eye on it...
Just on the last update (both cars on 2020.4.1), mine (black) has experienced a pretty aggressive phantom braking symptom twice. I did not pay attention to the situation specifics other than it was highway and in a slight curve left.
I have seen zero phantom braking since whatever revision we were on in November, and never previously was it this aggressive. Keeping an eye on it...
My phantom braking has only happened under one circumstance: nighttime underpass on the highway. It is a bitch. Driving along with no issue and come to the underpass and the car thinks it is going to hit a wall or something. Hard braking. I know that when I drive home after dark on the highway that I can't use AP. Eventually, they'll have enough images to be able to properly train the AI. Until that time...no AP at night.
Watch out. There are people on here that will call you a boldface liar or just a delusional uneducated driver when you say you had hard braking. They will tell you that only light regen braking is possible in these situations.
linux-works
Active Member
interesting - so I wonder if anyone is planning to add 'time of day' as a virtual sensor input, to be considered as part of the decision algorithms?
if its 'shadow time' of the day, the car could try to relax some of its strictness in object detection, if it can try to detect shadows.
its what a human would do, really. we don't ignore time of day and amount of light outside; it does affect our judgement. if we see something weird and the light is bad, we don't assume we saw an actual monster; the light plays tricks on us and we use better judgement.
these are the little lessons that have to be accumulated into the whole system, to try to achieve the safest self-driving system we can make.
OT: I wish we could work on this problem as a whole, meaning its each company trying to solve the same problem, but everyone working on it and sharing the results so we can get there faster. would be great if the world could get together and make it a common worldwide goal; to grow the system so that it can be used on all cars, as a service. in the middle of last century, john kennedy wanted the US to win the space race and he made it a point of national pride to get there, first. I wish we had a similar national visionary (not a corporate one). if we teamed up (world-wide common research problem) I'm sure we'd get there much faster and have a better solution than any one vendor could self-design.
if its 'shadow time' of the day, the car could try to relax some of its strictness in object detection, if it can try to detect shadows.
its what a human would do, really. we don't ignore time of day and amount of light outside; it does affect our judgement. if we see something weird and the light is bad, we don't assume we saw an actual monster; the light plays tricks on us and we use better judgement.
these are the little lessons that have to be accumulated into the whole system, to try to achieve the safest self-driving system we can make.
OT: I wish we could work on this problem as a whole, meaning its each company trying to solve the same problem, but everyone working on it and sharing the results so we can get there faster. would be great if the world could get together and make it a common worldwide goal; to grow the system so that it can be used on all cars, as a service. in the middle of last century, john kennedy wanted the US to win the space race and he made it a point of national pride to get there, first. I wish we had a similar national visionary (not a corporate one). if we teamed up (world-wide common research problem) I'm sure we'd get there much faster and have a better solution than any one vendor could self-design.
Kitfox
Member
Everybody has a difference of subjective opinion about braking. Some people call full regen braking "hard braking". Since nobody's expected to naturally know that "Losing more than 10 MPH/s is hard braking", and even if they did they're not going to be using a stopwatch and eyes on the speedometer, they're absolutely allowed to have their opinion.
Because of human nature and the way human beings are designed biologically, disconnects from input to action will also cause different impressions. You can be on TACC behind a "pulser" and have the passenger (and driver) think you are repeatedly braking even if there's positive power to the motor and different levels of acceleration at all times because of the constant changes the car makes to keep precise distance. (That is a different matter)
They might. Never seen an unexpected braking event from the car's point of view where the shadow of an overpass was actually sufficiently in view and range to look like it made a difference. That doesn't mean there isn't one. Only Tesla could say for sure, so report the things each time they happen. Squeaky hinge/wheel and grease and all.
Would be spiff. Corporate (and other) greed makes it nearly impossible in the current environment though
They've made some improvement, but it has a ways to go. They have now "Learn to wipe when I didn't know I should've" training. (Press the quick-wipe button when it's not wiping and you want it to and it will record that and send it for training if it can). They need to get a "Stop wiping! OMG!" training too. Or a way to visualize the front camera so maybe we can see why it's wiping. Maybe a bug splat.
GoergeJetson
Member
I'm still using AP even when it's dark, my phantom braking has also occurred in the early morning when I drive to work before sunrise. There are some lights on the interstate at each Express Lane sensor, kinda like a street light. I sometime get phantom braking when passing under those. So it's kinda the opposite of a daytime shadow, it's a lit area in the dark. I still like AP, even with it's inconsistent and frequent phantom braking, I just stay ready to tap the pedal when it happens, guess I'm just getting used to it.
Last edited:
Telstar
Member
I bought my car in the beginning of 2019 without any autopilot. Shortly after, I was given a 30-day free trial which I used on a long road trip. The only part of AP I really liked was the adaptive cruise control. It did the Phantom braking thing twice in 3 days on long straight country roads while using cruise. I didn't use cruise control for the rest of the trip, and I couldn't wait for it to expire so I can get my regular Cruise back. Freaked my wife out completely. She would have bought a model y but now she owns the new Lexus rx450h.
@Kitfox the detective can look at this
too bad the dashcam doesn't overlay any information like speed, but here we were going 136kmh on the left lane, and suddenly it jolted down to 110kmh by the time i glanced down at the speed.
often, cars in the adjacent lane would render as skipping in and out. its possible at that moment, it thinks the truck jumped into my lane (probably helped by its micro-movement left), and thus it applied the brakes harshly
i have more examples but it's a chore to extract all the files, scan for the right clip, upload, etc
too bad the dashcam doesn't overlay any information like speed, but here we were going 136kmh on the left lane, and suddenly it jolted down to 110kmh by the time i glanced down at the speed.
often, cars in the adjacent lane would render as skipping in and out. its possible at that moment, it thinks the truck jumped into my lane (probably helped by its micro-movement left), and thus it applied the brakes harshly
i have more examples but it's a chore to extract all the files, scan for the right clip, upload, etc
yes, the passenger gets the worst rendering of the incident. I think women (man, am I a nut or what to say this) react differently to a misbehaving inert object than a man does. Men are used to changing variables for some reason ;-)
Kitfox
Member
Three-second video, so it's a little short on context. Even checking frame velocity gets to be difficult in that short a view. I see a few things that would generate appropriate braking, and a few things that could generate inappropriate braking.
Most likely, assuming the inter-lane stripes at least match the most common international specs (~3m stripe, ~10m space), then the truck in the right lane is traveling about 34km/h (or more) slower than the Tesla average over those three seconds, so the Tesla is doing that whole speed adjustment based on the adjacent lane thing.
* That clip starts closer to the leading car and falls back. There's a chance that the radar detected a slowdown of the car in front of the leading car and was falling back for safety. Or, outside the clip, if that leading car was previously in the right lane and cut the Tesla off.
* Bridge crossing might be a road with a lower speed limit. Mislocation is a thing, a really crappy thing, and sucks, and should be reported regularly so Tesla can fix that {censored}. If the problem happens at that location any more in the future, report it as a bug with the voice controls if you connect to WiFi frequently
One thing they could definitely use improvement on is deceleration curve. If it's not an urgent maneuver, reduce the speed less urgently. The fact that people will think the car is "braking hard" even if it's just "accelerating less quickly" is a weird thing that may not have a good solution at all. But a more gradual change and ramp up of the force impressions could definitely help.
That's some good stuff right there!
I've noticed a lot of hesitancy recently by my M3. If I play close attention it seems to think that cars are crossing over the lane line into my lane when they are not actually doing so. It's often rendering cars as being over the lane line but they are often not even close to the line.
Not sure if my cameras are out of whack or what. I've only noticed this on 40.1.
I've had a few instances where the car slams on brakes for oncoming traffic - most notably around curves. Again - I think this is because the car things that the approaching vehicle is coming into my lane for a head-on collision.
I could be wrong.
Kitfox
Member
It's tricky, I'd think. The AP system has to be relatively conservative when it knows things so that it's not at fault legally as much as possible.
If given the choice of two systems, one that kills 125 people per ten billion miles, and one that kills 31 people per ten billion miles, with no other options right now and no other information is provided, most people would rather have only 31 people die than have 125 people die.*
If you tell the same people "Every ten billion miles it drives, Tesla's Autopilot will kill 31 people.", people will scream to outlaw autopilot, ban its use, conduct recalls, and so on, despite the fact that if humans were driving instead, 125 people would die. Even when aware of that fact, many people would rather have 125 people die from human actions than have 31 people die from machine actions, and would consider the machine unsafe even though it's a lot safer than the humans.
It's easier to accept the idea of blaming human beings for being stupid drivers than to accept the idea that a machine that is better than the humans overall is not perfect enough.
Everything that we can do to help improve the behavior is a net boon and many fewer bad happenings over time. The more you can understand how the car thinks, the better you can be aware of what it will do and the safer you will be.
* Bonus
When working with really big numbers and averages, they can often be difficult to comprehend. "125 people per ten billion miles" will cause a lot of folks to say "Well, I don't drive even close to ten billion miles! And that's a lot, so that's, what, ten years of everybody in the US driving?"
In reality, that's around a day of everybody in the US driving. So 125 dead people per day, unless we get every car in the US on AP, in which case only 31 a day, and getting better. But never better fast enough for people to not fret.
If given the choice of two systems, one that kills 125 people per ten billion miles, and one that kills 31 people per ten billion miles, with no other options right now and no other information is provided, most people would rather have only 31 people die than have 125 people die.*
If you tell the same people "Every ten billion miles it drives, Tesla's Autopilot will kill 31 people.", people will scream to outlaw autopilot, ban its use, conduct recalls, and so on, despite the fact that if humans were driving instead, 125 people would die. Even when aware of that fact, many people would rather have 125 people die from human actions than have 31 people die from machine actions, and would consider the machine unsafe even though it's a lot safer than the humans.
It's easier to accept the idea of blaming human beings for being stupid drivers than to accept the idea that a machine that is better than the humans overall is not perfect enough.
Everything that we can do to help improve the behavior is a net boon and many fewer bad happenings over time. The more you can understand how the car thinks, the better you can be aware of what it will do and the safer you will be.
* Bonus
When working with really big numbers and averages, they can often be difficult to comprehend. "125 people per ten billion miles" will cause a lot of folks to say "Well, I don't drive even close to ten billion miles! And that's a lot, so that's, what, ten years of everybody in the US driving?"
In reality, that's around a day of everybody in the US driving. So 125 dead people per day, unless we get every car in the US on AP, in which case only 31 a day, and getting better. But never better fast enough for people to not fret.
ElectricIAC
Good-Natured Rascal
I’d like to turn off that adjacent lane horse manure. It’s been nothing but problems for me.
NOA works fine for me on my commute and I use it every day so.. (with auto lane changing turned off) It's been well worth it for me since I have a long commute.
DopeGhoti
Active Member
I very much like the adjacent lane change adaptation. I see far too many idiots that jump out into the lane ahead of me, and I'm glad to not have to stand on the brakes when they do so like I did before. I can understand how some mightn't like it, though, so sure, I'm all for a toggle so that folks that don't like it can turn it off.
It took me a while, but I actually rather like auto-lane-change without affirmative confirmation now. You do still need to bless lane changes- it will not start a maneuver until it's done a hands-on-wheel check. This coupled with the wibbly-wheel alert rather than chimes makes NoA quite seamless and relaxing. I would, though, prefer the car to not start signaling until after it's gotten permission to make the change.
It took me a while, but I actually rather like auto-lane-change without affirmative confirmation now. You do still need to bless lane changes- it will not start a maneuver until it's done a hands-on-wheel check. This coupled with the wibbly-wheel alert rather than chimes makes NoA quite seamless and relaxing. I would, though, prefer the car to not start signaling until after it's gotten permission to make the change.
Similar threads
