I'm pretty sure it's the tie downs flaps from the curtains ..flapping around in the wind. I don't get any problems beside container trucks. But any trailer with flapping tie downs can cause my phantom braking.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Cruise control no longer safe
- Thread starter Giri99
- Start date
I’d find it very useful if people could better describe the road type and conditions when reporting phantom braking or similar issues, as well as the follow distance they have TACC set to.
A lot seem to be occurring on single carriage A and B roads, or in town, which I don’t personally believe are appropriate places to be using cruise, irrespective of automated driving tech.
My experience so far with autopilot on motorways, with follow distance set to 5 is that it’s no worse than the adaptive cruise in my previous car (a Jaguar XF). No problems with passing trucks, and only a couple of slow downs but no panic braking. I don’t use cruise when it’s busy and only on motorway or dual carriageways.
A lot seem to be occurring on single carriage A and B roads, or in town, which I don’t personally believe are appropriate places to be using cruise, irrespective of automated driving tech.
My experience so far with autopilot on motorways, with follow distance set to 5 is that it’s no worse than the adaptive cruise in my previous car (a Jaguar XF). No problems with passing trucks, and only a couple of slow downs but no panic braking. I don’t use cruise when it’s busy and only on motorway or dual carriageways.
Lucky you. I don't count single carriageway instances. My two most severe experiences were on the M6 and the M40 and I also have follow at 5 spaces. M6 was perhaps understandable in as much as plenty of traffic but on the M40 case there were no bridges or traffic in the middle lane and fortunately nothing behind.
Glan gluaisne
Active Member
My experience of the car swerving/braking has been without TACC engaged, on a stretch of the A30. It does it fairly consistently, every time there is a large vehicle waiting to come out of a side road that runs at about a 45° angle to the main road. It hasn't ever done it when there's just been cars waiting at that junction. There's another junction, right on a bend, that's triggered it a couple of times, too. I'm pretty sure it's related to light conditions, as well as vehicle type, as it's never done it in rain, nor in bright sunshine.
I'm not sure about the straps, but I'm 100% with you about it being curtain siders that are the most predictable at causing phantom braking, especially when there is no lane changing involved.
Also 100% that there is some underlying fundamental and dangerous behaviour that is not getting resolved.
I reckon I can predict 30-50% of phantom braking on motorways. My current most likely trigger is a full length artic, white with no signage or writing, slightly loose / flapping siding.
The other related issue I find very suspicious links back to green's early (and subsequent) posts showing clouds of radar reflections from a variety of objects that are not actually in front of the radar path.
My opinion is that this is secondary radar reflections, especially when linked to over head signs and passing under bridges.
Each trigger has very specific causes / conditions, and the only one I so far disagree with is shadows. I do not think the shadow is relevant, but the object casting the shadow could be.
Objects jumping around and suddenly changing type / orientation is a concern if the visualisation we see actually has any correlation with what the car bases decisions on. The other day I pulled up along side quite a high wall and the visualisation turned it into an intermittent large lorry!
I've followed so many threads about old 'fake' fsd and new 'real' fsd, 4D rewrites etc, and the one thing I was hoping for was that visualisations would become stable / non jittery. I do not understand how a safe system can be prepared to habitually and consistently be in a state of uncertainty / flux about what it is seeing. By the time you are along side a truck, your system logic should tell you that truck can't step directly sideways into your path.
My understanding was that by introducing stitched video from all sources and tracking objects in that combined feed a lot of random impossibilities would be eliminated.
I suppose there is a conflict between knowing certainty based on analysis and prediction compared with an ability to respond to the unpredictable. At the beginning, a NN can believe anything, including the completely impossible. It seems that even with a lot of training, very thin threads of the impossible still remain.
Last edited:
Glan gluaisne
Active Member
If it's a radar related issue, then the easy fix is to switch to using monopulse. This was the fix used to get terrain following radar working very reliably decades ago. By just gating the return acceptance window, then most of the spurious multipath type interference just gets gated out. I can't honestly believe that the system that Tesla are using doesn't already do this, given the complex target environment it's operating within.
I wouldn't take anything for granted!
Does gating reduce resolution in some way? Is it possible with all radar system designs? Does it introduce blind spots that would need multiple radar to eradicate?
Maybe Tesla is pushing their radar to extremes that would make gating limit performance too much for their application?
Last edited:
Glan gluaisne
Active Member
All monopulse gating does is eliminate returns that occur a set time after the transmit pulse. This limits range, but also cuts out all multipath reflections, as they have a time of flight that's too long to be true returns. The system works by transmitting a single pulse, then only processing returns that occur within a set window of time after that pulse. Once the gate closes, the next pulses is transmitted. It won't reduce the clutter from very short range multipath, but will reduce a lot of the stuff coming back from multiple longer total path length reflections. Combined with looking at the Doppler shift on the returns, and knowing the vehicle speed over the ground accurately, allows a fair degree of information to be collated, even given the wide beam of the radar, and the fact that it isn't, AFAIK, capable of scanning that beam.
@Glan gluaisne, I've wondered about flapping straps / buckles etc. and dismissed them as too few / too small to have an impact, but is that actually the case? Presumably wavelengths are broadly the same now as in all radar applications and susceptible to the 'window' effect.
Glan gluaisne
Active Member
If I had to guess, then I'd say that it may well be the Doppler shift from the faster moving flapping bits that's fooling the radar return. Doppler shift is used to determine relative velocity between objects sensed and the car, so something flapping might be interpreted as having a higher relative velocity, perhaps towards the car, then it really has. If so, that might explain why curtainsiders seem to be a particular problem.
Perhaps it is a radar issue. Despite saying many times that the current hardware is sufficient to provide full autonomy (rather than just FSD, which is anything but full autonomy) Tesla have seemingly decided that the current radar is inadequate. They are apparently sourcing a “4D” radar sensor from Arbe Robotics.
Tesla is adding a new '4D' radar with twice the range for self-driving - Electrek
MrBadger
Badger out
Do you have any thoughts for this one from @Pagemakers ? Lane change bug Great video showing the phantom right/left movement.
My guess fwii is that if you could see the multi-camera feed that went along with that event, it would be fairly clear where the 'misunderstanding ' / misinterpretation comes from. My thoughts are also inline with the issue being more pronounced when the truck is further away with an 'empty' lane imbetween.
I assume that at the present time, all images are processed as 'flat' with uniform resolution. It looks like the issue relates to confusion in the interpretation of perspective from trying to make 3D assumptions based on images from different cameras with different and changing perspective. The sudden jumps are, I would suggest, where you switch between camera views. The biggest perspective errors will come when the images being tracked are furthest away. [edit: not just furthest away, but towards the extremes of the camera's field of view]. The gradual movement of the truck side to side could be perspective compensation error within a single camera's view, and the sudden jump when you switch from one camera to another. You (well at least I) can often see visualisions of vehicles showing a false curved trajectory, I assume as they approach certain extremes of image boundaries.
I hoped (hope) that by applying more sophisticated image processing to obtain a much better 360 view (that would tie in with EM's suggestion that birds eye might be a future feature) you would then be feeding a far more unified / homogenous / integrated view into the FSD / AP system. The jitters / jumps and other anomalies could all be largely due to the non-linearities of taking flat 2D images and trying to interpret them as 360 / wrapped images.
[edit: I guess this perspective stuff would be critical in city driving where you have to process 360 imagery with a high degree of stability / confidence. The 'hand-over' between camera's and perspective issues would be a nightmare to deal with, so may be that was a large part of the re-write. You can probably make far safer assumptions / predictions about your environment driving along a motorway than in other situations, yet phantom braking could be an example of even motorway driving needing to work with a single integrated 360 view.]
Last edited:
....... relating to phantom braking specifically....?
Given the relative jitteryness and uncertainty of (at least the visualisation version) visual images as interpreted by the car, I don't see how emergency type responses can be based on them - certainly not alone. I haven't been driving the MS much recently, but in the past I have seen reflections (esp of brake lights) on wet road surfaces show up as multiple random traffic cones so you wouldn't want to risk responding to that version. On the otherhand, if radar or ultrasonics pick something up, there is maybe a better case for taking immediate action.
Last edited:
MrBadger
Badger out
My concern is that we have quite possibly had sensor fusion resulting in BEV (birds eye view) since around this time last year - and according to green a brief, possibly shadow only, more sophisticated (plaidnet) version around 2020.8 iirc, . What is not clear is what, if any, identification and labeling is being done on the BEV for this scenario. As it sounds like you have noticed for yourself, the city streets FSD beta is still not robust, at least not as much as I had expected, with some visualisations. Dancing and phantom wireframes still exist relatively frequently, and I'm not talking about road layouts, but specifically objects or phantom objects. Whilst I have convinced myself that city streets/FSD beta is still very much a completely different domain to that of highways driving (indicated by the different UI visualisation), I remain concerned that for highways (ie regular TACC and Autosteer when used in the documented domains) are not going to be solved in their entirety by the 4D/BEV. I would actually go as far as saying that, some of the more recent behaviour that we are seeing is actually as a result of the use of BEV and not a will be fixed by BEV. The missing link being labelling across the 4 dimensions - and I hope that will go some way to resolve objects transporting or morphing over time.
I had a weird one a few months back. Pulling away with other traffic from having been stuck at a level crossing for several minutes was immediately in front of us, a cyclist and immediately in front of that a transit type van. As we started cornering to the right, then back to the left (S bend), the visualisation clearly showed the cyclist smoothly overtaking the van. To the point that I was impressed by how smooth the visualisation was. Unfortunately, as I looked up, the reality was the van actually pulling away from the cyclist, yet the visualisation had it the other way around with the cyclist appearing to race off into the distance.
And here we see the challenge of 'computer / AI interpretaion'. Is it a motor bike? Cyclist? Electric bike?.... Which scenario is possible / most likely?
The perspective issue became very evident to me a couple of months back when stopped at a new set of traffic lights with small 'eye level repeaters' similar to on the continent. The visualisation turned these into multiple distant traffic lights, all with their own red line across the road way.
Since in the UK at least it is legal / acceptable to drive with only one eye, there are some limits to depth perception based on a flat image that humans at least can work around.
I have no idea of the correct terminlogy, but I can see a conceptual image that is actually produced by differences between consecutive frames. It's not really just a visual image, but there is a visual element to it, along with a certainty / probability / persistence element. I wonder if the NN is trained purely based on visual evidence from camera frames, or if there is any other component derived from analysis of multiple consecutive (?stitched BEV?) images?
Last edited:
I think the categories of automated driving need revision in light of Tesla’s performance...there should be a -1 level where you need to be more aware than if you were driving unassisted
I've had similar issues with the phantom braking - I've also had issues with the car trying to leave the motorway/dual carriageway at certain junctions, this can be quite disconcerting!
Sorry to take this back a few posts. But in my experience it's not just curtain siders. I have had the same effect passing trailers towed by cars or pickups with ties holding down vehicles or small excavators and one which was covered by a tarpaulin held down by ties.
My last trip South I purposely avoided all of these and had no phantom braking for 500 miles.
But I still get caught out by bridges and I can't figure that one out.
My last trip South I purposely avoided all of these and had no phantom braking for 500 miles.
But I still get caught out by bridges and I can't figure that one out.
