David W Virtue
Member
Overhead caution lights that weren’t lit still stop my M3, so I’m assuming the feature detects shapes of traffic lights but not the colors. Yet.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Elon: "Feature complete for full self driving this year"
- Thread starter dckiwi
- Start date
mikes_fsd
Banned
It detects both, the current release notes (and when you turn the feature on) state that this feature will begin conservatively by stopping at all signals.
Just like it will currently stop at a signalized railroad crossing.
You know how I know that City NOA/ 100% Feature Complete FSD wont occur this year? Because Smart Summon (at 3mph) STILL will occassionally take my car up and onto a curb. In broad daylight. In yesterdays case, with zero other cars within 60 feet of it in the parking lot at Sams club. When there was a completely clear driving lane adjacent to the curb.
I still think there are blind spots to the cameras and or sensors, that need to be addressed via additonal sensors/hardware before the current cars can truly achieve the full goals that Elon has promised.
I still think there are blind spots to the cameras and or sensors, that need to be addressed via additonal sensors/hardware before the current cars can truly achieve the full goals that Elon has promised.
willow_hiller
Well-Known Member
It's worth watching Karpathy's latest talk about the Autopilot rewrite:
Smart Summon is currently run using the older Autopilot code that is looking for curbs independently between camera views. Once the rewrite is deployed, the neural network will be able to synthesize all of the camera views into an overhead view of the entire intersection.
It is a stretch that feature complete FSD will be done this year, but Tesla's pace of innovation is faster than what most people realize.
diplomat33
Average guy who loves autonomous vehicles
In case, you all missed this, update 2020.16.2 appears to bring some improvements to the visualizations including animated pedestrians.
As you probably noticed, he is projecting a video in front of the car's front cameras. He is not actually driving. So, the cameras are simply reacting to the video it sees. So it's like a simulation.
As you probably noticed, he is projecting a video in front of the car's front cameras. He is not actually driving. So, the cameras are simply reacting to the video it sees. So it's like a simulation.
Thanks!
If interested, I’ve attached a couple photos of where the incident occurred. Went back today to Sam’s to pickup couple of things. When the smart summon failure occurred the other day, I was in the parking spot directly in front of where my car is in the photo. At the time of curb incident, there were no cars nearby at all. I marked the spot in yellow where the car went up onto the sidewalk/ curb.
Attachments
Last edited:
Their pace if using known techniques isn't faster though. These are all standard technique in the industry, I mean its good that they are using it, they should. But its not some "innovation".
I hope that depth from monocular which was popularized by Google Brain in CVPR 2018 and is further iterated by many others to this day and is currently a standard in CV
and seg views to BEV model which almost everyone has a model variation on isn't what you are putting all your hopes and dreams on?
TLCTimFL
Member
I hope Tesla smooths out the steering during FSD. I was riding east on a 2 lane 40 mph country rd at around 45 mph letting the M3 do the driving on a perfectly clear, sunny day. (I'm still at 2.5. I tried 3 times to set an appointment for the 3.0 retrofit. Each time they contacted me back and said they didn't have the part yet. But they finally told me on try 3 that they would contact me when the part arrived.)
So back to my 2 lane rd...My M3 came upon an area that had a new housing development on my right in which they needed a center turn lane for the westbound traffic. So my lane went toward the right (to make room for the center turn lane), then straight to get past the entrance, left and then straight to get lined back up with the original road. I felt like a pinball being thrown around the car while the M3 jerked the wheel in each change of direction. The computer has to see what's going on along the road and ease each turn of the wheel.
So back to my 2 lane rd...My M3 came upon an area that had a new housing development on my right in which they needed a center turn lane for the westbound traffic. So my lane went toward the right (to make room for the center turn lane), then straight to get past the entrance, left and then straight to get lined back up with the original road. I felt like a pinball being thrown around the car while the M3 jerked the wheel in each change of direction. The computer has to see what's going on along the road and ease each turn of the wheel.
Soda Popinski
Member
The caveat in your example is you're running HW2.5, which isn't running the FSD "preview." As such, it's not suppose to be used on country roads. I'm not even sure how well a HW3 with FSD preview would handle it, but at least with the intersection code running, they can't claim it's only for use on divided highways with no intersection.
Knightshade
Well-Known Member
So you used the system someplace it's explicitly not intended to be used, and it didn't work right, and you're blaming the system?
I think Tesla was really crystal clear in what to expect with this feature. That one is to expect that occasionally there will be a false positives or false negatives. I look forwards to trying it out............
..........But, this feature in itself isn't what the concern is.
The concern was there is a stacking of features that can all cause false braking.
To those of us that understand how these things work it's not a big deal. If I encounter a lot of false positives with this latest feature I'll likely simply turn this feature off.] END QUOTE...
I'm about to write something with tongue-in-cheek, but in reality, if the false braking issue is as bad as some have reported, here is
a suggestion. In our state, when you allow a new driver to use your car you must attack a bumper sticker to the rear of the car
that says: "NOVICE DRIVER".Sounds as though we need to attach such a sticker if we intend on using newer versions
of AP since it behaves like a novice driver.
Perhaps the bumper sticker will alert those who insist on tailgating or not paying attention while behind us...to pass or beware.
Google Image Result for https://cdn.compliancesigns.com/media/NH/truck-transportation/1000/Driver-Traffic-Safety-Label-NHE-15853_1000.gif
The sense that the car has little or no awareness that a good driver would get from reading the road ahead is the main thing that makes me doubt viable FSD any time soon. It may be getting better in the US where it sounds as though the vehicle does slow ahead of stop signs and intersections, but in the UK I see little if any of that.
In one recent short journey, I did get the impression that speed approaching some bends was better controlled, but in general little sense of speed control ahead of potential hazard zones.
As an FSD / HW3 owner, I am of course eagerly awaiting the day when suddenly my MSR grows up and passes Drivers' Ed, but in the mean time I stuck suffering my 'teenager' battling with a protracted and uncertain 'learning' phase.
Watching this video was not very encouraging for me. It felt like 'The King's / Emperor's New Clothes' (emperor’s new clothes, the | meaning of emperor’s new clothes, the in Longman Dictionary of Contemporary English | LDOCE). Very few clear statements. Lots of vague generalities. Lots of potential psuedo-techy terms. +1 fluff, -1 fact. I would almost prefer to hear nothing and get used to the way my car actually is rather than keep being reminded that it should be better (soon). That said, of course it is good to be told that 'we are working very hard on loads of good stuff', and I have high hopes for the release of the re-write but at the same time am bracing myself for something less than hoped for.
Last edited:
When we only have glimpses of what is going on with FSD development it is hard to pull together a coherent picture. From the Karpathy videos (starting with Autonomy Day) and the recent third row interview with the ex Tesla employee it seems like HW3 gives the processing power to take the full fidelity inputs from the cameras (and other sensors) to produce the 360/pseudo Lidar 3D model and 3D labelling. If this is correct then the result of the rewrite will be the first time that FSD perception will be working to the full capability of all the sensors. It also seems that 'memory' will exist both in the dynamic 3D model (it should be aware of things that were previously visible but are currently obscured - like a cyclist obscured by a truck) and in the map (there is a traffic light around the corner, even though it is not yet visible).
Like you I have high hopes that the rewrite will provide the platform for resolving the 'teenager' and awareness issues you refer to. Also like you I am not expecting it to be perfect initially
It will be interesting to see if the car 'remembers' its surroundings when it is restarted - so that it remembers the position of curbs etc that it can no longer see but which were visible as it drove to that position prior to being parked/switched off.
Knightshade
Well-Known Member
Again, it's not designed or intended to be used at all in places with stop signs and intersections (outside of the very early beta update for US cars only just very recently launched- mainly for data gathering rather than useful functionality)- says so right in the owners manual.
Everyone is still basically running the "base" 2.x code that's years old and fundamentally not going to work for FSD- and Tesla has openly said so.
Basing any thoughts on "it'll never work" using how the CURRENT code behaves is a waste of everyones time- this isn't the code that's going to get there and not meant to be at this point.
Hence the fundamental re-write that will (allegedly) view the world in a completely different way.
Now, is THAT code going to be good enough? Nobody knows. But we KNOW the current one isn't, so projecting ahead based on it continues to make no sense.
And if they really do manage to get this done, then they might even be able to do the 360 bird's-eye for parking! After all, even without a front fender camera the cameras will have previously seen what's there and know how far the car has since moved.
It certainly feels as though with the current approach, you can see the visualisation fighting between what each camera currently thinks it can see, so the new approach to image processing may address that.
Yes, reading the road ahead. Stop signs and traffic lights - no sign of them being handled in UK, so I'm not referring to those.
Even though my comments about reading the road ahead 'as any good driver attempts to do' do not relate to city driving, I still maintain that the development to date seems to have resulted in a rather miopic, single plane view of the road ahead with the car having no problem changing its mind multiple times about what it thinks it sees. The concept of the car possibly constructing and maintaining a persistent view of its current situation based on history as well as current view is interesting.
I agree that discussion about the 'old approach' Tesla has now given up on is somewhat academic and hopefully not predictive of things to come. And you only learn by making mistakes!
My speculation is more based on the process and approach that delivered the current version as well as the apparent lack of clarity over how bug reports and driver interventions have contributed.
In the Third Row interview, I was left unconvinced about what happens.... if anything.... in response to Bug Report &/or driver intervention over AP. The impression from the interview is that every driver intervention results in a report back to Tesla, but fell short of indicating if such data actually resulted in specific changes being implemented. What I have read elsewhere suggests that 'bug reports' don't even leave the vehicle and no evidence of interventions still routinely being taken into account by Tesla.
The question posed in the interview about feedback from other countries and how that was handled also interested me as I get the distinct impression having driven regularly in the USA, France and UK that AP tries to apply driving rules and lane positioning that is clearly based on US road layouts and conventions.
Making the re-write fundamentally relevant to each country where the cars are sold is also an important evolution that needs to happen.
Last edited:
Knightshade
Well-Known Member
Bug reports absolutely do not leave the vehicle except when you open a service ticket for your car and the service center can access them remotely to see what he problem with your car is.
Greentheonly has confirmed this among tons of SC folks telling customers same.
As for disengagements he's written quite a bit about those too, and the specific campaigns Tesla uses to gather data.
tldr version is Tesla has grossly overstated what "shadow" mode does- though it does do some things.
Similar threads
