I believe FSD has gone almost as far as it can giving the British road system coupled with UK driving regulations. The FSD roadmap has always been biased to USA driving with long freeways, large roundabouts and junctions. Although I have FSD I'm not holding out much hope for any major improvements for the UK. I hope I'm wrong.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Tesla to launch Full Self Driving beta to select drivers next week
- Thread starter timpharrison
- Start date
Username does not check out.
JupiterMan
Active Member
In theory I think this is actually quite straight forward, as if the car knows which lane it is in and where it is heading, it will from experience be able to label the appropriate traffic light(s) that apply to your lane. once this is mapped (by you or any other Tesla driver) it will remember it. In many ways FSD will be better than humans in this regard as it will have the benefit of the fleets experience, and going to a new junction which would fluster a human should not fluster FSD.
Whether theory works in practice we shall see, I'm an optimist on all this!
If after the rewrite it can do 5 laps of the M25 unscathed and without breaking any laws, it will be a major achievement.
There is a good chance that at some point it will use auto park rather than lane change
I wonder what NOA will make of the Dartford crossing. So many lanes to choose from...
There is a good chance that at some point it will use auto park rather than lane change
I wonder what NOA will make of the Dartford crossing. So many lanes to choose from...
rincewind
Member
This particular junction doesn't have straight lanes - both left and right directions curve off to the left and right, but the lights are not placed to reflect that. I admit it's a strange one, but not unique I'm sure. As you say, the machine learning may help here.
I think its Teslas definition of FSD thats Musk is referring to not the definition that says the car will drive itself at level 4 or 5 autonomy.
The feature complete argument is not yet there, nobody has seen capabilities to merge into traffic as an example at say a junctionm on to a motorway. There was an assessment somewhere that went through them all and Tesla were missing lots of "features".
I think the best we can expect is a more robust capability doing what it currently does very variably. So fewer phantom braking incidents, better lane discipline, better reading of speed limits and stop signs etc. But thats still not going to let you turn across oncoming traffic either at traffic lights or on a normal road, still not going to let you turn out of a side road etc.
As for mapping. I've never been a subscriber to that being the answer. I think its an aid and may improve speed, but the real world doesn't stay stationary and you don't want to be the first on the scene where its different, either contraflow in roadworks when you're on the wrong side of the road or a new road layout or there's a broken down lorry. And somebody has to be first to each of those scenarios - and if you can cope with the change, why did you need the map?
The feature complete argument is not yet there, nobody has seen capabilities to merge into traffic as an example at say a junctionm on to a motorway. There was an assessment somewhere that went through them all and Tesla were missing lots of "features".
I think the best we can expect is a more robust capability doing what it currently does very variably. So fewer phantom braking incidents, better lane discipline, better reading of speed limits and stop signs etc. But thats still not going to let you turn across oncoming traffic either at traffic lights or on a normal road, still not going to let you turn out of a side road etc.
As for mapping. I've never been a subscriber to that being the answer. I think its an aid and may improve speed, but the real world doesn't stay stationary and you don't want to be the first on the scene where its different, either contraflow in roadworks when you're on the wrong side of the road or a new road layout or there's a broken down lorry. And somebody has to be first to each of those scenarios - and if you can cope with the change, why did you need the map?
ML generally doesn't work by remembering special cases - that it the old style of programming. ML is more about finding the hints which are always present but not often noticed - so if this junction looks a bit like lots of other junctions it can be deduced, otherwise it will always be a problem. The deduction can of course factor in any mapping data, and inferences about lane division where there are no lines (based on all the cases where there are lines, or poor quality lines).
JupiterMan
Active Member
As I understand it, Tesla does maps for traffic lights - for instance FSD will be aware and will flag a traffic light as oncoming even if it is round a blind corner.... of course it will only ever know colour of light once it has a visual.
But likewise, it will pick up temporary traffic lights in road works, so it is using both maps and visuals here.
I would not be surprised if it also maps other features over time, such as zebra crossings and speed limit signs so it can plan ahead.
Where Tesla's approach differs to some others is that it does not do high resolution mapping, which as I understand it actually plots via the map the cars trajectory and relies on that map heavily. I agree with you that this approach has many problems should things change, and can reasonably only be expected to ever work within a confined geo-fenced area.
isn't an exact opposite also referred to as a negative? there you go again being all negative!
HenryFarkas
Google "Henry Farkas Tesla"
I don't have any insider knowledge, but if I were writing the FSD system, I'd use the neural network that connects every Tesla back to the Tesla FSD cloud to obtain information about each of the little byways all over the world. I'd get the network to collect information about how previous Teslas have driven down each and every byway, either driven by the driver or by the car, as long as each drive didn't end in a crash. Then, I'd send said information about a particular locality back down to each Tesla that happens to be in that locality. Each car gets advanced knowledge about its own area of operation, but it also gets general information about how to deal with new roads not previously known to the network. If the car can't decide what to do, it gives control back to the driver and records what the driver does and how well that works. Eventually, the network gets so good at driving that it never needs to give control back to the driver. It also knows to avoid little roads that it can't understand. Usually, there's an alternate route.
Glan gluaisne
Active Member
I have a feeling that the Google/Waymo system does pretty much this, with the 3D map that every vehicle uses being augmented all the time by live updates from all the other connected vehicles. My guess is that doing this for all the fixed objects reduces the real time processing requirement a bit, by simplifying some of the initial object classification stuff. Whether that makes a worthwhile improvement to the overall capability or not I don't know, I'm just guessing that anything that helps to reduce the real time processing requirement is probably a good thing.
That wouldn't handle unusual circumstances, other drivers, parked cars, lorries being towed backwards, a trailer full of traffic lights (infinite TL spawn), pedestrians, pedestrians dressed as cones etc. I'm pretty sure that it will already pretty well handle driving on more or less any deserted road.
What will be interested in is what's referred to as the Dojo computer, able to take video recorded data and train the model, by understanding is the training is currently more limited to still images.
HenryFarkas
Google "Henry Farkas Tesla"
Remember, Elon didn't say the rewrite would be actual FSD. He says that his car can go from home to work with nearly no intervention. He didn't say that he never had to intervene. It's still a work in process, and those of us, like myself, who purchased FSD, knew or should have known, that we are the people who are training the neural net. It's our job to stay alert and intervene properly as soon as intervention is needed. If you weren't OK with that, you shouldn't have bought the FSD. I'm not associated with Tesla in any way, but I am considering buying some stock. So you can interpret my remarks with that in mind.
Sparkeur
Member
Asking for a bit much there aren't you? That's what, 500 miles? It would have to charge you up for that.
Mad max mode on the M25?
Glan gluaisne
Active Member
I bet more people buy a Tesla because of features like the performance than ever buy it because they believe that it will drive itself. Be interesting to do a UK Tesla purchaser's poll to find the most popular reasons for choosing a Tesla. I know that range and performance were probably the top two on my list. I'd guess the supercharger network and some of the tech in the car probably ranks quite highly, too.
Similar threads
