I sometimes wonder if a future hardware version will (or should) include an articulated telephoto camera, with dedicated sub-networks for aiming the camera and interpreting the telephoto view. Map data is far too often wrong for FSD not to override it based on prima facie evidence, such as road signs or painted road markings.
I guess I don't know what I expected, exactly. I guess I figured somehow there might be a bit more variation with an NN. It's not exactly the same input. Could be daytime, could be nighttime, could be cloudy, could be sunny, could be raining. The relevant pieces are the same of course.
I wonder if yellow light times are coded into map data? If not, maybe they should be? Yellow light times are typically correlated with speed limits (faster limits = longer yellow lights), but there is no real standardization, and red-light-traps do exist. There have been attempts to propose formal standards, but in practice it's still sort of the Wild West out there. At the same time, there are prima facie rules in some states (most states?) that one should stop at a yellow light if it is possible and safe to do so. (If there's a car right behind you, it's arguably not safe to do so.) Some states prohibit accelerating to make it through yellow lights. It's possible that Tesla is taking the lowest common denominator approach that would be the most legal in most states, though in some cases it may compromise safety. I hope this improves as time goes on.
I would like it to not tap the brakes, then proceed, when the left-hand-turn light goes yellow and my light stays green (I have video of this, which I absolutely will not post). How on earth did the miraculous NN learn to do that completely incorrect behavior?
I don't want to be a Negative Nancy, but in the most optimistic sense of speaking, the chances of this are zero with current hardware.
I don't know how they input navigation directions into the end-the-end model, there are different ways you can do this. But one day they'll probably need some multi-modal thing that can handle language tokens like "Turn left on 10th street", and that will make it possible to grok the roadway street signs. And you can give it plain directions like, "park near the west entrance to Costco".
I do expect that they have some sort of Venus Flytrap approach where it takes at least 2 disengagements at a specific location for them to start paying attention to it. [A venus flytrap only snaps shut if it detects two successive movements in nearly the same spot, to prevent random events like raindrops triggering it.] But agreed, subsequent capture of non-FSD driving at these locations should be very helpful.
Another weird behavior I've noticed with 12.3.3 is that it will occasionally get stuck at 5-7mph OVER the set "Maximum" limit (e.g. if the speed limit changes) and won't slow down even if the maximum limit is adjusted with the scroll wheel; it requires a full disengagement and re-engagement in that case to get it to slow down. Seems like a C++ bug more than a neural net error.
Yes, there's a "Use HOV Lanes" setting in the Nav preferences. Although I do wish they had an intermediate "Adaptive" setting, using the seat weight sensors to determine occupancy, and using HOV lanes if and only if it determines enough people are in the car. Relatedly, it would be fun to see a "Watt-Hours per Passenger Mile" metric displayed in the Energy stats.
On the contrary; deep-learning models are highly nonlinear. infinitesimal changes in input can result in vastly different changes in output. The important thing is that the behavior should be stable "when it counts", e.g. for not curbing wheels. And in the real world, photon noise in the cameras is enough to make a considerable difference in how the car navigates an otherwise-identical scenario if encountered twice in a row.
As I understand it, they implemented no-rolling-stops in v12 by carefully curating their training set to avoid rolling-stop examples, not by hand-coding. It's not clear if this approach is pragmatically scalable to every such parameter, though.
My assumption has been that they are using tons of simulated stops, in order to program the fundamental stop-sign behavior and achieve the non-human-like profile. (Though it obviously still needs work).
It’s one of those short entrance/exits where as the interstate rises into an overpass you enter/exit at the top. This type of enter exit lane is short and you don’t have much time to exchange with incoming traffic. FSD was signaling and preparing to get into the lane and exit, but then decided to slow and allow incoming traffic to get in front of it. A no no for this kind of exit as the traffic merging on is suppose to enter behind you, while they are still speeding up. I’ve used this particular exit multiple times a week for over 10 years and never ever missed it.
Did you read my very next post? I basically say the same. Here is link. Also slight correction. It is Summon. Summons is a legal term.
teslamotorsclub.com
You got STUedtoday and it was not a Dislike but a Laughy.
Is this a first for the bot?
Link as proof to post.
teslamotorsclub.com
