-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
But it didn't turn there and if you follow the path indicator it never intended to turn there. However, it shouldn't have started the turn signal before passing that street. Turn signals starting too early or too late have been pointed out in other videos. Just another bug to work out.
Daniel in SD
(supervised)
He runs a red light, then almost runs a red light (it turns green a split second before the crosswalk), then turns right from the incorrect lane, then posts it to the internet.
In the past a lot of algorithms have been used to track objects, often from different sensors sometimes called object fusion. Most of these algorithms were analytical and very feature engineered. Recently there has been some progress of using machine learning ruch as RNN to improve tracking and lately even transformers to select the most likely candidates for objects to track. See for example https://arxiv.org/pdf/2005.12872.pdf
Imo what Tesla are doing right now is moving away from an object tracking algorithm to an occupancy grid prediction. The neural network will just ”learn” that the past positions of objects are relevant for the future positions of objects and thus ”track them”. Then there is no maximum number of objects it can learn to track, but it can only predict one object per pixel in its output space(if output space is for example 200x200pixels that is a large number). But I guess in theory it can track even more. Imo if you feed the neural network some video streams of thousands of people moving around, it will learn to track them. Can see some value for using simulation to test this.
I think it is likely that Tesla some time in the future will switch from a CNN/RNN style neural network to a transformer style network. There they might have to actually specify the number of objects that should be tracked. Not sure, the field is pretty new and lots of progress is happening.
Testing Autopilot/FSD's ability to drive blind - 12:49 - AIDRIVR
Quick run to Albertson's to pickup some groceries - 2020.44.10.2 - 24 Nov 2020 - 4:14 - James Locke
not very good. There was one particular moment at 24:30 where the car’s pathing had it turning right right into a barrier and the driver saved it at the last second. No drivable space at all there and nothing that would fool it, just a complete **** up by AP. Scary as *sugar*. Hopefully the v5 of he beta coming out that Elon touted addresses a lot of these issues, because as is, it’s not close to being ready for a general release.
powertoold
Active Member
Yup, I don't think so either, still around 5-8 months to go IMO. But then again, Elon has said that we have trouble estimating exponential improvement timelines. But but then again, Elon also said we'd have Tesla robotaxis by now
‘it’s making right and left turns with the pillar and repeater cameras AND the front facing wide view blocked? That isn’t good. Default if it can’t see should be to sit it’s ass at the intersection and blare warnings that the driver needs to take over.
One thing is pretty clear to me thus far. There are basically two groups of people when it comes to FSD. Those that think it has to be perfect (or very close to it) for a wider release and those that feel that it's already substantially better than what we currently have.Yup, I don't think so either, still around 5-8 months to go IMO. But then again, Elon has said that we have trouble estimating exponential improvement timelines. But but then again, Elon also said we'd have Tesla robotaxis by now
I'm squarely in the latter camp. I use autopilot every time I get in the car. I use it on every road that it will allow me to use it. The current release of FSD Beta seems MUCH better and safer than what I'm using today.
I'm guessing that the ones who fit in the first camp only use AP on divided highways.
Nothing wrong with either view so don't take my post for something that it's not.
I still think we will get a broader FSD release before year-end. Maybe very early January, 2021. But I don't see it being much later than that.
powertoold
Active Member
I don't think he blocked the wide angle camera enough. Also, it's likely that Tesla enables FSD in these situations for the beta but will likely have a contingency feature for multiple blocked cameras in the future.
It's great that the car remembers the 3D environment even when cameras are blocked. That's the main point of the video I think.
It isn’t substantially better, though. It is more capable of doing more things, but at a potential decided loss of safety. At 24:30 of the video in DC, it ignored the road to try to make a turn that wasn’t there, right into a barrier and a curb and grass beyond those two things. I’ve never seen regular AP lose its lane keeping like that. Since regular AP can’t make turns, that isn’t something that you have to really worry about, whereas it is with the FSD beta since regular AP isn’t capable of making turns, while FSD beta is.
At 4:15 he clearly says he is blocking the wide and narrow forward cameras (the pillar and repeater cameras are still blocked). At about 5:08 it approaches an intersection, you can see the right and left of the intersection dissolve as the normal forward facing camera can’t see *sugar*, and then it makes a left anyways. That is dangerous as *sugar* and should NEVER, EVER happen. If it can’t see down the streets of the perpendicular street at an intersection, it should NEVER, EVER proceed.
Similar threads
