It's not an entirely new program. All the logic is the same and it differs only in the amount of input data. We will see.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Autopilot goof. Good thing I was paying attention :)
- Thread starter bcsteeve
- Start date
-
- Tags
- autopiliot
Knightshade
Well-Known Member
Nope.
Go back and listen to the Q3 2018 investor call
Andrej Karpathy --Tesla Director of AI & Autopilot Vision said:
It's not just able to run existing programs with more input- it's able to run a different, much larger and more complex programs.
The ability to do that is why they bothered to develop the hardware and why it's required for FSD.
Autopilot wants to stay in the middle of a lane, so when the center line disappears it thinks there is a very wide lane and moves to the middle. Then when a line reappears in the middle it wants to pick a line to be in. Seems to be working as designed right now, as it is not currently meant to be used on a non-divided road. It correctly picked a non-occupied lane (just happened to be the wrong one though). When FSD is available, I'm sure it will be able to tell if it is a divided highway or not. You are pushing the limits of AutoPilot, but also using it on exactly the kind of road it is currently not supporting.
This is not true in the sense that when the road is wide, autopilot will assume there's more than one lane with implied lane boundaries.
While you cannot know this for sure on other cars, in my example it's very clearly seen on the augmented view:
wrongway-trim - Streamable and wrongway-right-trim - Streamable
So Autopilot picks a lane, which is perfectly acceptable for what AP is supposed to be used for currently. Either lane is good.
Wow crazy! i think some of these are just going to need to be reported to tesla (voice command: Report a problem with ...). I did that over and over again for a curve in my area close to an intersection where the markings disappear (nothing as bad though as the example in the video shown)
somewhere in the last 6-12 months they fixed it (I don't think they suddenly improved the AI, I bet someone at Tesla can manually review and place some overrides on the map??)
How would AP1 handle that same situation one wonders?
I think it's great that people are pushing the envelope and thus sending Tesla disengagement reports to allow them to construct a better AP.
BTW, the problem is that freeways and surface roads are constructed differently and AP is currently biased towards freeways.
In freeway on ramp and merging situations, you can indeed have a lane that widens up and thus finding the center can make sense.
But on surface roads, lanes rarely widen up, and your algorithm should be to hug an appropriate side lane marker.
Regardless, this is just evidence that Tesla hasn't released an AP that is reliably capable of driving on surface roads yet, something we all knew.
BTW, the problem is that freeways and surface roads are constructed differently and AP is currently biased towards freeways.
In freeway on ramp and merging situations, you can indeed have a lane that widens up and thus finding the center can make sense.
But on surface roads, lanes rarely widen up, and your algorithm should be to hug an appropriate side lane marker.
Regardless, this is just evidence that Tesla hasn't released an AP that is reliably capable of driving on surface roads yet, something we all knew.
No. It's more complicated than that. not only it knows of the yellow doubleline while it sees it. it knows this is a two lane road with opposite direction of travel from the maps. So there's just a bit of a split brain issue.
it is a split brain.
I can observe autopilot operations directly.
But you can do a simple test:
setup wifi metering for your car. Then take a drive somewhere make sure it's longer than 2km. return in wifi range and go into park. you'll see very small wifi upload spike. Like under 100kb.
Then do the same drive, but engage autopilot and then disengage in any way. return to wifi and observe ther's no huge upload.
Also you can refer to my twitter thread that documents TEsla data collection: green on Twitter
Since the software for FSD doesn't exist yet, what Musk is really saying, in his typical over-enthusiastic way, is that it would be possible to run the software if it existed. Until the software exists, nobody knows what hardware will actually be needed. I am of the opinion that for optimal full self-driving, the car will need more sensors than it has now, as well as a more powerful computer than HW 2.5.
My Model 3 drives wonderfully well on surface roads, within the criteria of Level 2 (hands on the wheel, eyes on the road) autonomy. I have to take over much more often on surface roads than on highways, and I have to make the turns, but EAP still makes the car safer than without it. It is very much capable of Level 2 driving on surface roads.
Iti trivially obvious that Tesla has not yet released software cabable of driving without human intervention. ("Yet" is an important word in that sentence.)
this does not do anything useful for autopilot bugs. also only service center typically looks at this data when you have other complaints, it's not even is transferred to tesla mothership otherwise.
Kant.Ing
Member
OBX John
Autonomous Driving Enthusiast
It's not a "good thing" to pay attention while on autopilot.
It's expected that you pay attention at all times while driving. Which you are still doing, just with autopilot assistance.
I fear that as autopilot continues to improve, this attitude that it's driving the car instead of you will get worse.
It's expected that you pay attention at all times while driving. Which you are still doing, just with autopilot assistance.
I fear that as autopilot continues to improve, this attitude that it's driving the car instead of you will get worse.
CameronB
Member
This isn't what I see on local two lane roads (one lane each direction). AP display shows only one lane (the one I am in). It ignores the oncoming lane(s) completely. Therefore "it just sees lanes" doesn't seem to be accurate. Now obviously it is using lane lines to (largely) determine which lane(s) are available to you and which to ignore (i.e. which are oncoming lanes). Though I've seen times when the lane markings were completely non-existent and it still worked. This happened to a local road by my house a few months ago when they re-paved the road but hadn't got to painting the replacement lines on it. So there was about a ½ mile stretch that had zero painted lines dividing the two opposing lanes. AP worked fine. It didn't try to center itself in one giant lane.
Also just to add my two cents to the highly debated "should people use AP on city/surface street" argument: Obviously people need to be extra vigilant when doing so. Hyper-Vigilant might be a better term. But if you are vigilant and are aware of the limitations off-highway, it can be useful at times.
Despite what the manual says, Tesla has subtly encouraged off-highway usage. Case in point: for some wider multi-lane local roads that function as "expressways", 9 months or even 6 months ago, AP, if activated, would not show you additional lanes right or left and would not allow you to change lanes (while on AP) in these situations. There was some block from Tesla that told your car you can't do that here.
Then a certain software update happened (don't remember which one) and all of a sudden, AP could "see" the adjacent lanes on the expressway that it wouldn't previously...and it would happily let you change lanes on the expressway in the same way as you could on a highway/freeway. I'd argue that Tesla didn't allow their fleet to change lanes on expressways by accident and therefore it was a subtle way to get people to do lane changes on the expressways and give Tesla more data about how that was working.
Similar threads
