Pink Duck
Active Member
112 pages of that report can be found at https://unece.org/sites/default/files/2023-06/ECE-TRANS-WP.29-2023-86e.pdf
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
UK FSD Discussion
- Thread starter Buckminster
- Start date
JupiterMan
Active Member
My definition would be : an automous car at level 4, which by that definition would involve geofencing and likely some weather restrictions as well.
At level 4 the cars would be autonomous and could be operated as RoboTaxi's, just like Waymo is now, however over time, the Geofencing for tesla would not be because of mapping perimeters (like Waymo) but because of legal jurisdictions, i.e. within one state that permits it.
MrBadger
Badger out
Autonomous, as per SAE definition, is Level 3, 4, 5
The big difference between L3 and L4, is that L3 requires a legal driver to be at the wheel (and be in a position to take over in a controlled manner when vehicle leaves its ODD, maybe speed limit or type of road), where as L4 (or L5) does not require a legal driver to be in the vehicle, and thus does not necessarily require that vehicle to have a steering wheel.
Difference between L4 and L5 is that L5 requires the vehicle to drive in any situation that a human may reasonably expect to be able to drive, but L4 (and L3) will have a operational designed domain where the vehicle should be able to drive, and outside that would either require a driver or (if no legal driver) not be able to operate.
Important to differentiate a L3 controlled handover with uncontrolled/panic take over that we experience now. L3-5 should not suffer an uncontrolled/panic takeover and safely failover.
The big difference between L3 and L4, is that L3 requires a legal driver to be at the wheel (and be in a position to take over in a controlled manner when vehicle leaves its ODD, maybe speed limit or type of road), where as L4 (or L5) does not require a legal driver to be in the vehicle, and thus does not necessarily require that vehicle to have a steering wheel.
Difference between L4 and L5 is that L5 requires the vehicle to drive in any situation that a human may reasonably expect to be able to drive, but L4 (and L3) will have a operational designed domain where the vehicle should be able to drive, and outside that would either require a driver or (if no legal driver) not be able to operate.
Important to differentiate a L3 controlled handover with uncontrolled/panic take over that we experience now. L3-5 should not suffer an uncontrolled/panic takeover and safely failover.
Pink Duck
Active Member
Automated Driving Bill cleared the House of Lords yesterday and so on to Royal Assent stage it goes (when Rishi next weekly meets with Charles).
Plus Elon posting that FSD 12.4 Supervised will remove the steering nag in the US - probably on request of NHTSA to rely exclusively on cabin camera.
Plus Elon posting that FSD 12.4 Supervised will remove the steering nag in the US - probably on request of NHTSA to rely exclusively on cabin camera.
FSD refuses to engage/work if the camera is covered.
Ah right, did not know that. It doesn't seem to stop Autosteer working over here (which is pretty much all we've got).. it also disables attentiveness checks (eyes looking off the road too long). I've heard it does increase the steering input nag but not confirmed.
Yeah, you can still cover the camera and use Autosteer here, at least for now. (But FSD requires the camera to work, if your car has it. It, also, requires the single-pull activation method.)
To be fair if cabin camera monitoring effectively replaces the steering wheel nag entirely in the next update and finally delivers a form of handsfree driving, I'm all for it!
I'd love to see this and them to catch up with what BlueCruise already offers in the country.
Yeah, required for FSD. And NHTSA seems to be upset that users have to opt-in, and can turn it off at will. So it is possible Tesla will make that standard for Autosteer as well.
But if they do, they really need to make an alternate way to activate TACC by itself without having to switch configurations/profiles.
it’s a bit geeky but I recon this explains AI and the differing views on whether it’s going to ‘exponentially improve’ (which always seemed a daft assumption to me but seems to be a legitimate argument), through to the more pessimistic view that it will plateau before reaching optimum performance.
It’s not about self driving as such, but the principles and arguments on AI I think are the crux as to whether Tesla will deliver genuine L3/4 let alone L5, and as it’s not about Tesla, or self driving, there’s no edge or bias.
#discuss
It’s not about self driving as such, but the principles and arguments on AI I think are the crux as to whether Tesla will deliver genuine L3/4 let alone L5, and as it’s not about Tesla, or self driving, there’s no edge or bias.
#discuss
Pink Duck
Active Member
I agree with that assessment in general. Tesla have moved on to solving the edge cases of their AI driven self-driving approach, but recently stated just how few percentage of human driving clips there were to do so. So they are/have moved on to simulated life-like edge cases for positive/negative training.
It did occur to me that as individual human drivers we may never encounter or have to deal with these cases ourselves, so with time the AI systems will have an edge on humans by eliminating the reaction time to 'novel' situations. That’s where the power of large datasets can win out over our own lifetime driving experiences.
Ultimately things will plateau depending on the human time/resource spent on curating the input data sets, pending generalised AI where the AI itself can self-refine somehow.
It did occur to me that as individual human drivers we may never encounter or have to deal with these cases ourselves, so with time the AI systems will have an edge on humans by eliminating the reaction time to 'novel' situations. That’s where the power of large datasets can win out over our own lifetime driving experiences.
Ultimately things will plateau depending on the human time/resource spent on curating the input data sets, pending generalised AI where the AI itself can self-refine somehow.
Absolutely will be diminishing returns. If you have footage of people driving a piece of road and train on 10 of those clips, it might not be all that good. Now go and train it with 1,000 clips of the same and it'll have proved greatly. Train it on 100,000 clips and it'll be better but maybe only marginally. I mean numbers here made up but you'll get the point. The more data you train on, the more it's going to cost so decision will be where is it good enough because the additional cost for minimal gain might not be worth it.
Think we might have a ways to go yet though and you'd assume that when we get to this goof enough, the car will be vastly better than a human driver anyway.
It's a shame I didn't have a Tesla in my earlier years. I could have shared them some footage when I half drove off a bridge, another time when I tried to wade in a Corsa and ended up floating. There's the time I skillfully avoided a jack knifed lorry. Another time I went up on two wheels to dodge a home delivery van. Maybe the crowning moment was when I was a passenger with my friend driving a hired Ford Ka, saw a humpback bridge and yelled "Jump!" so he did. I can tell you they don't have a 50 / 50 weight distribution as they don't fly through the air like the General Lee did.
I could go on, I've always been a very good driver
That would be awesome if that's true as would bring it up to BlueCruise levels but I guess work on any road Autopilot already does rather than geofenced roads?
I wouldn't get too excited, these kinds of amendments pass through GRVA all the time. It's unlikely the first text will be voted through, so that means more data gathering/redrafting for subsequent meetings, plus once it gets through that stage it has to go to the full WP29 which can take several months even if they don't make any changes.
It's a slow bureaucratic process. If we see anything practical in under 12 months I'd be surprised. More likely a couple of years.
It's a slow bureaucratic process. If we see anything practical in under 12 months I'd be surprised. More likely a couple of years.
Of course. I know they need to consider safety but there's also the productivity issue they need to consider. Self driving cars will then mean things like self driving taxi's, self driving lorry's and so on. If the US has these, the costs to their business of moving things around the US drops and they become more competitive. If we still need humans to do that same work, we are missing out economically.
We focus on what it means for personal cars but it's the business implications of this technology that will make a real difference.
