-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Major Phantom Braking Issues
- Thread starter LucLamma
- Start date
not only that... imagine someone behind you and they must be thinking you are check-braking them... suddenly braking on a wide open road with no traffic...
They happen at the same circumstance. The road continues, there is a dip which has road which is invisible to the camera and the road continues after the dip and becomes visible again to the camera at a longer distance. But if it looked for vanishing lines (as I suspect it does) it saw a discontinuity / mismatch. I bet that's the origin of the problem, it detects what it thinks is contradictory data, and going at a high speed it's programmed to slow down as it doesn't know what to do until it resolves into a continuous lane line it can estimate to the horizon.
I'm not making an excuse for it, but a diagnosis. I've seen this same effect in YT videos a year ago, and it was often at undulations.
Since it happens because of the road geometry then presumably crowdsourced data could find those locations of consistent failures, obtain better training data to overcome the problem, or mark these road areas where the logic should be changed to not slow with such a discontinuity.
It is claimed that the FSDbeta code is less reliant on lane lines than the current highway AP code (which is running this), so maybe it will be less susceptible when the long awaited 'merge' happens. That's still a ways off for non FSD drivers for standard AP. I bet they also didn't fix it before because management (I.e. Elon) thought the FSDbeta merge was right around the corner but it obviously wasn't. Tesla has a remarkably small autopilot development team (only a few dozen actual developers/ML scientists) for the size of the task they have to support (many regions, many models if you include all the sensor & computer variants). Elon chronically understaffs and overworks. With Tesla's revenue and tremendous profit margins they should not be as stingy on a key technology R&D area.
sleepydoc
Well-Known Member
That’s roughly what my car was doing for quite some time. It was usually about 10 mph before I would catch it and it might not have braked quite as hard but pretty close. When I had the gall to complain and criticize Tesla for it I was told all cars do it and criticized for being a hater and having unreasonable expectations. All I can say to @pcopeland is it’s not acceptable and we all should expect better of any car.
To be fair, from what I hear everyone else not driving a Tesla has a better TACC so their car would not bang into his Tesla
to be fair, using Radar would *not* trigger such phantom braking ... because mirages, dipping roads and other optical illusions aren't fooling a radar signal. "vision only" was the worst cost cutting from Tesla's side and software delete the radar input in older cars just the icing on the cake.To be fair, from what I hear everyone else not driving a Tesla has a better TACC so their car would not bang into his Tesla
sleepydoc
Well-Known Member
I’ve seen other people post that their car seems not to phantom brake when there’s a car behind them. That’s a difficult hypothesis to test, but raises other questions if true.To be fair, from what I hear everyone else not driving a Tesla has a better TACC so their car would not bang into his Tesla
I can’t make any observations on it one way or another.
stopcrazypp
Well-Known Member
Or rather if there is no car behind, phantom braking doesn't come with the possibility of causing a rear end accident. I have long suggested this is the way it may be done. Basically you can err to allow more false positives when there is no one behind, given slowing down doesn't put you in any additional risk.
I also find little PB on crowded SoCal freeways. It must be frustrating to get PB on wide open highways with little traffic where it should be very 'easy' for even a dumb system.
I have the feeling that the PB issue might be an unexpected emergent property of a complex software system that they haven't been able to engineer out without hurting something else. They're moving more towards neural networks for the driving policy and control system, which are also opaque and incomprehensible from a logical point of view, but one can adjust the behavior by changing the weighting of various examples in the dataset and also giving positive and negative reinforcement. The problem is that you need a huge database of correct and well curated proper driving behaviors to train against. That's a more complex problem as it deals with time series and contextual information than the datasets they have acquired for visual classification (autolabeler) which is now a solved technology. That requires only following academic published literature and then doing a great job of it in practice.
But there is very little written in open publications about training driving policies as it's more commercially focused. And difficult.
GlmnAlyAirCar
Active Member
zoomer0056
Active Member
just for context from my experience... if I follow a car there is no PB. all the PBs i had were on wide open roads with no car in front of me.
that being said ... if "vision only" needs a reference point like a car ahead to "latch onto" as visual reference - and otherwise can get confused.... not a great selling point and probably one of the reasons Tesla is moving back to radar for their next gen cars.
that being said ... if "vision only" needs a reference point like a car ahead to "latch onto" as visual reference - and otherwise can get confused.... not a great selling point and probably one of the reasons Tesla is moving back to radar for their next gen cars.
I think this is my biggest pain point after the fact that my 2018 M3LR has radar that didn't have this problem and it was disabled. I don't like that Tesla disabled the radar and pushed Tesla Vision while it still has these issues, but if I could at least set the car to a non-traffic-aware, "dumb" cruise control, I'd at least have a temporary remedy while they work on it. Growing up along I-94 in North Dakota we often joked the road was so flat, straight, and empty you could set cruise control and take a nap as you drove from one end of the state to the other. It should be the easiest case for self driving that there is
+1I think this is my biggest pain point after the fact that my 2018 M3LR has radar that didn't have this problem and it was disabled. I don't like that Tesla disabled the radar and pushed Tesla Vision while it still has these issues, but if I could at least set the car to a non-traffic-aware, "dumb" cruise control, I'd at least have a temporary remedy while they work on it. Growing up along I-94 in North Dakota we often joked the road was so flat, straight, and empty you could set cruise control and take a nap as you drove from one end of the state to the other. It should be the easiest case for self driving that there is
my biggest issue with radar-deleted TACC are the PB events on empty highways... *exactly* where I usually prefer to use TACC and where I barely had any PB events before my radar unit was deleted
FastEddieB
Member
FL trip completed and safe back home In TN.
2,815 miles altogether. Still no phantom braking on all sorts of roads and conditions.
Not saying it’s not a very real problem for some, and I sympathize. But still favoring my hypothesis that some combinations of car and software are more problematic than others - and maybe we just got lucky.
2,815 miles altogether. Still no phantom braking on all sorts of roads and conditions.
Not saying it’s not a very real problem for some, and I sympathize. But still favoring my hypothesis that some combinations of car and software are more problematic than others - and maybe we just got lucky.
Similar threads
