Silicon Desert
Well-Known Member
Well, I guess we would do well together. got my license the same year and no accidents yet in over 700,000 miles. Now that I said that, my car will be totaled tomorrow
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
FSD rewrite will go out on Oct 20 to limited beta
- Thread starter diplomat33
- Start date
Are there academic papers on traffic intersection detail inference on-the-fly, or is this all Tesla-proprietary?
I see papers like "Navigating Occluded Intersections with Autonomous Vehicles using Deep Reinforcement Learning",
or "Anytime Lane-Level Intersection Estimation Based on Trajectories of Other Traffic Participants". From a quick
scan, I can't tell what kind of a priori intersection geometry is used in the former, or, for the latter, how this could
be useful for Tesla without other drivers at the intersection.
Color me impressed just from the North Carolina video.
Last edited:
powertoold
Active Member
Nope, this is all Tesla. I believe Tesla is able to do magical things with their NN because they own the entire stack. They can use the car's sensors (steering position, accelerometers, odometers, GPS, etc.) to measure all sorts of road characteristics derived from vision. Because they are able to control every aspect of their sensor positioning and angles, focal lengths, etc. etc., along with their huge fleet, they can label large amounts of data using "future" labels to label past images. That's how they're able to achieve other cars' positions and speed despite the cameras being partially blocked by a front car here:
https://twitter.com/teslaownersSV/status/1320098691680665600
Edit: who am I kidding? I'm probably not even scratching the surface of what Tesla is doing to achieve mindboggling NN predictions.
Last edited:
Daniel in SD
(supervised)
I had two accidents within a year about 250,000 miles ago. Both times rear ended while stopped at a light! One other when I was 17 (I was the third car in a chain reaction, crappy drum brakes.).
Basically I've got to go the rest of my life accident free to get to the 500,000 mile average. I suspect the true average is closer to 250k miles.
An important point is that not only does FSD have to avoid at-fault accidents it also has to avoid accidents where it might not legally be at fault. Many collisions are the result of two people not driving optimally.
Silicon Desert
Well-Known Member
Glad to hear your average is higher. That's saying a lot considering you live in San Diego. Wait a second, I'm not picking on the drivers there. I go to SD a lot and I am more worried when driving there. Some of the street and turns are a bit strange to me. Then putting the heavy traffic and drivers in a hurry into the mix, my blood pressure goes up.
I guess if I lived there, it would be different and I would be used to it, but as a visitor, I keep a very watchful eye on things.
Daniel in SD
(supervised)
Haha. I learned to drive in Los Angeles, San Diego is easy mode. Insurance rates are half what they are in LA.Glad to hear your average is higher. That's saying a lot considering you live in San Diego. Wait a second, I'm not picking on the drivers there. I go to SD a lot and I am more worried when driving there. Some of the street and turns are a bit strange to me. Then putting the heavy traffic and drivers in a hurry into the mix, my blood pressure goes up.
Silicon Desert
Well-Known Member
Oh, I see... "LA", some say stands for LOT of ACCIDENTS. Rarely go there. Frankly I think there are lot of accidents in any big city.
That's probably why you are more alert these days. Now back to the code re-write after my digressing from the topic. Get that FSD to work in LA."City Streets" is what Tesla calls the FSD beta that people are using. So "City Streets" was driving in the rain. ("Navigate on Autopilot" is what does the driving on freeways/highways.)
This guy is driving on some brand new roads and where brand new, non-functional, traffic lights have been installed:
Nope, most of the people have had very few followers, at least until they started releasing footage.
It seems to me like they have been targeting leaders of the various Tesla Owner's clubs and asking them to share their experiences with the beta City Streets feature.
Huh, this guy says that you hit the camera icon and it sends Tesla the info to check out the clip for problems. I’ve seen other guys do it too. But I haven’t seen Brandon hit that icon yet, I don’t think. Maybe he wasn’t told to hit the icon?
Saw this tweet. Does the system actually learn on the fly? I though all the data was uploaded to the mothership and changes would be made there. Now it seems it is learning as it goes and probably uploads all it has learned..
https://twitter.com/alex_avoigt/status/1320489656396427264?s=21
https://twitter.com/alex_avoigt/status/1320489656396427264?s=21
It's pretty clear (and you obviously have sources) that Tesla does not consider maps that they use as "HD maps" despite them containing quite detailed info that some might consider to be "hd maps" based on the amount of extra finegrained detail they have over "regular navigation maps". But, again, Tesla does not consider them "hd maps".
With that out of the way, you somehow seem to interpret my retweet in question as a statement that if Tesla is using hd maps, then their approach does not work without said hd maps, OR because Tesla approach does seem to be working everywhere in US - they therefore have HD maps of entire United States (which you don't believe yourself and therefore conclude that Tesla does not use hd maps based on this fact alone).
Because nothing like that was stated by me (or even tweets I retweeted) and because I find this notion ridiculous, I am trying to understand how come what I said could be misinterpreted in such a horrible way.
There are plenty of simple roundabouts especially if they're lower traffic. Here's a couple examples of visual barriers to help drivers focus and yield to those coming from the left. Also, the visual barriers help drivers notice a roundabout is coming up and not to just go straight.
Kietzke & Neil in Reno
Highway 50 & USA Parkway near Carson City
FSD beta handles a gravel hilly road covered with leaves like a champ in north Carolina. so impressive
Tesla Full Self-Driving Beta Demo: Hilly Gravel Road in North Carolina [VIDEO] - TeslaNorth.com
Tesla Full Self-Driving Beta Demo: Hilly Gravel Road in North Carolina [VIDEO] - TeslaNorth.com
What the neural network sees, thinks and what is visualized is not the same thing. If you look at the latest karpathy presentation often the output of the neural network is blurry, ie it thinks it is a 25% chance there is a barrier at position x,y.
A fun mental exercise, it might think there are either 2 or 3 lanes with equal likelihood, what will the neural network output? Answer it will make 3(1+2) 50% opacity lane markers making it look like 4 lanes. What should be visualized in this case? I think it just visualizes everything above a certain threshhold, say 50%, see for example some examples of cars disappearing and appearing in the visualization in the latest fsd clips being posted. But the neural networks sees and thinks about everything below this threshold also.
That’s pretty good. Would be really impressed if it allowed a car coming in the other direction to pass
Last edited:
You posted a video from 2014 and say “just announced”?
Similar threads
- Question
