Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

Autopilot consistency and accuracy...

This site may earn commission on affiliate links.
I really wish Tesla would utilize the cabin selfie cam for driver monitoring, similar to comma.ai and super cruise, but as an augmentation to wheel torque. Most of the situations described above, other than phantom braking, would naturally lead to an attentive driver putting their hands back on the steering wheel to be ready for evasive action (blinding sun, sharp turns, erratic drivers, etc.), but during long stretches where nothing is happening, it would be so much more relaxing to not have to torque the wheel. And yes, I understand that to do it right, you’d need to add some LED IR cabin lighting to allow it to see in the dark and through sun glasses (a simple add-on accessory could accomplish this until Tesla builds it in). I also don’t know if the current selfie cam has enough resolution, but listening to George Hotz on multiple podcasts, it sound like driver monitoring with a cell phone camera was the easiest thing in the world for comma.ai. Given the position of the model 3 camera, so close to the driver’s face, I bet it could be done with the current hardware.

The other thing that I think could be improved with AP, is to add a dead pedal to the right of the accelerator. I always hover my foot above the accelerator, because I’m worried about phantom braking, or so I can close the gap during stop and go. A right sided dead pedal would be a nice cure for the hoverfoot.
 
Last edited:
  • Like
Reactions: Siciliano
The camera isn't remotely in the correct place for this (apart from being the wrong type, lack of lighting, etc that you mention)

Look at where the camera is for this on Caddys (or the Mach E for that matter).

As to comma.ai- apart from the fact they dodge a ton of regulation that Tesla would have to deal with- the phone they're using has a 16 megapixel front camera and their kit comes with... guess what?


Comma ai website said:
Infrared LEDs for night time driver monitoring

The camera in the Model 3, IIRC, is 1.2 megapixels. 15 times lower resolution.
 
  • Funny
Reactions: Daniel in SD
The camera isn't remotely in the correct place for this (apart from being the wrong type, lack of lighting, etc that you mention)

Look at where the camera is for this on Caddys (or the Mach E for that matter).

As to comma.ai- apart from the fact they dodge a ton of regulation that Tesla would have to deal with- the phone they're using has a 16 megapixel front camera and their kit comes with... guess what?




The camera in the Model 3, IIRC, is 1.2 megapixels. 15 times lower resolution.

Yeah Elon needs to swallow his pride and just integrate LIDAR into the system. The more redundancy the better.

Like I've said, I do think it can be done with cameras alone, but they probably should be a human eye equivalent resolution (576 MP), and the software needs to be powered by artificial general intelligence. Without that, actual FSD (w/ZERO human intervention) is just a pipe dream.

This is why I didn't buy FSD...while we will probably have small cameras in the near future capable of 576MP, we're about 15-20 years away from AGI, which is a mandatory component for lvl 5 autonomy.
 
  • Like
Reactions: TLLMRRJ
Yeah but so what? Do you really WANT that "level 5"? At 100 fatalities a day caused by humans, you look at Tesla's record and realize that it's already vastly hugely safer than said humans. And yet that's still considered "shaky" !?. Maybe the biggest problem is that the mere handful of accidents with autopilot, versus 35,000 FATALITIES - not just accidents, caused by humans every year in the US alone, are enough for these humans to complain. People want the impossible. A completely autonomous vehicle has to deal with HUMANS doing crazy stuff, and resolve the liability issues. Hell, we have 40% of the population that's eager to accept fairy tales over anything rationally obvious, like the fact that planetary temperatures are rapidly rising, and that it has a lot to do with atmospheric pollution. No matter the sensors, the social and psychological problems are probably impossible to solve at this time. Meaning vehicle automation is best used by humans at level 3 or maybe 4 for the foreseeable future. And I'm very happy having that, using it to improve safety and reduce fatigue, and watching it develop.

The "more redundancy the better" in a robotics system is not correct. Past a certain point it's well known that additional layers introduce more complexity and more integration problems than they solve. People who talk about buying it "when FSD is ready" are welcome not to have it, but they have no idea what they're talking about. That's like the people who rode horse-drawn buggies for years because cars were not ready. Or they can have someone design their dream cuisinart that's bristling with a hundred thousand dollars in sensors and so much self-testing code that it's impossible to maintain.

As to the torque debate, hand on wheel is more conducive to quick corrective reaction than having the human stare ahead with his hands far from the wheel. Maybe you also need high rez sensors to monitor where the feet are, and so on and so on, what kind of music is playing, if the driver is day-dreaming, and of what ...
.
 
Last edited:
I'm very confident and it greatly helps on the long highway trips I take quite frequently. It's very consistent for me and keeping 1 hand resting on the wheel works well and completely rids the warning from popping up. There's always the cheater "weight" ;) .... but annoying to have something on the wheel - unless of course it's small and has quick-release.

Also, as stated above, it's awesome for stop-and-go traffic.
I use an ankle weight but frequently still keep my hands on the wheel, but my arthritis doesn't like "hands on" for too long
 
  • Like
Reactions: XLR82XS
My M3 autopilot still crushes the breaks when it approaches a highway overpass that is casting a strong shadow.

I guess it sees the solid black and thinks it's going to hit a wall. Not great for driving around NJ where overpasses are extremely common.
 
I am a relative newbie, I got my Model X in Dec last year. I have a 2020 model x long range + with FSD. I agree with most of those above. My limited experience indicates as stated "Most of time" it works fine, but when it fails you do not have a lot of time to recover before you would get into an acident.

I have found that it has a LOT of issues when you drive through construction areas, emergency areas or places where they have changed lane markings for some reason.

For those reasons, I agree Never fully trust FSD. Use it but do not abuse it.

When I go on trips, I always stop for a break at least once every 2 hours or so I can at least get out of car and relax for a few min. I have found it allows me to be more attentive while driving long distances.
 
I love my M3... but when I initially engage AP on the highway the default cruise speed sets to 90 and accelerates FAST while I am trying to panic scroll back down to 70-80. Very annoying and wastes energy that I may need to conserve. How is the default speed determined?
 
So how confident are you guys in the autopilot?? Confident enough to take your eyes off the road? I haven't really done that (much) yet, but I have used it while eating a sandwich. It's pretty amazing, basically the future, now. FSD will be interesting when it's fully fleshed out...

2017 S 100D. 43k miles. Several long trips (> 3k mike round trip) and one xcountry... 10k mi round trip.

I use FSD almost all the time on the highway, occasionally off highway to see how it behaves. On the highway I turn off auto Nav when the car persists in changing lanes when the look ahead (further than car can see) clearly says you shouldn’t...

I Force the car to change lanes left when I see cars in the on ramp that will be a problem if I don’t. The car always manages the situation, but not without scaring the crap out of the other vehicles when the car finally “sees” them.

intake full control when the car insists on slowing to try and squeeze in between two cars to my right (to line up for an exit) when a slight tap on the gas to get ahead of the train of cars to the right is clearly the better option. I get it.. the car does not want to use acceleration in such cases. And truth be told, when I let the car do its thing, the cars to the right make way... the car just has a better opinion of other drivers than I do.

while it has improved, it is still damn scary passing 18 wheelers. Why won’t Tesla allow the car to drift a little left (or right in the inside lane) to give just a little more space?

and finally, I take control when there suddenly is debris on the road. Skunk, retread, pothole... the car does not seem to notice and will happily drive over them.

But I really like letting the car drive. There have been at least 5 situations where the car probably did better than I would.. or at least, I hope I would have reacted in time. So on balance, an amazing piece of technology. Don’t leave home without it!
 
Hope you don't get into an accident where the front wheels get torqued one way or the other. In addition to the damage to the car, you may end up with two broken wrists. My understanding is that Formula 1 (and probably other race car drivers) will let go of the steering "wheel" if they anticipate a crash.
 
I have a Model Y long range with the fsd computer. When it is doing the steering it seems to hug the left edge of the lane when going around a curve. Once it clipped a cement curb which was a center divider. On a trip I just got back from it would occasionally cross the center line. In one instance the tires were on the rumble strip in the center of the road. The problem happens when the road curves in either direction.

Anyone else having this problem?
 
Hope you don't get into an accident where the front wheels get torqued one way or the other. In addition to the damage to the car, you may end up with two broken wrists. My understanding is that Formula 1 (and probably other race car drivers) will let go of the steering "wheel" if they anticipate a crash.
I‘ll second that. Not sure how prevalent the problem is, but I’ve heard of people losing thumbs from airbag deployments. Resting your wrists on the center of the wheel could be even worse, if you can’t get them out of the way in time.

Air Bags Are Hazardous To Your Thumbs @ Top Speed
 
@JerryDoma - Never had what you describe. Not sure why some cars act different from others. But we all have our peeves, and you learn to anticipate. In my case I don't like how it centers in the combined space where the line widens before an entry, i.e. where another car will be merging in. I would prefer if it stayed more to the left. But give it some time. You'll see that every update is different, it's a work in progress, things gradually improve.
 
I always pay attention but I am confident in it's abilities and know it's limits. I notice that around 75mph or above, it has a little "issue" at curvy areas of highway, in that it leans too much on to the opposite side for my liking. If it curves right the car is too far to the left where it's almost touching the lane lines. Same when it's curving left. Now, at around 70, it does it pretty good even on curvy highways.

Rarely I get phantom breaking and sometimes slows down too much when i'm switching lanes. It's good enough that I can look elsewhere for longer periods on time. However it's important to pay attention because I've had to dodge items on the road.
 
I love my M3... but when I initially engage AP on the highway the default cruise speed sets to 90 and accelerates FAST while I am trying to panic scroll back down to 70-80. Very annoying and wastes energy that I may need to conserve. How is the default speed determined?
go to your settings. you can set your default to be relative (like +5 over speed limit) or absolute (like 90 mph! are you kidding me?!!!)
 
I would trust the FSD/AP but so far I am unimpressed with the vizualisation on the screen.
They lag for about 1s or 1m precision.
Each time a car is overtaking me from usual blindspot, when I check between actual visual and screen representation, the car is not where it is shown on the screen.

For instance, on screen i will see the car still behind me while in reality it is alreay 1m in. Representation adjust when (I believe) side cams catch it and then the car move faster to its actual position. This is annoying.
I also had the experience of stopping the car before a stop line and it was shown in the screen as if half the car passed the stop line.

So, as long as this is not fix (is there a way I can calibrate it ??) I woudn't trust it.
 
My wife was driving a bit during a road trip, and we never thought about how her profile had never been initialized. So she had no NOA, and effectively had "base autopilot", no automatic lane changes, and she was mostly just using Cruise. THAT's where we had a few "braking" events. For instance when she wanted to change lanes a bit close to the car in the adjacent lane. I've always suspected that unexpected slowdowns are mostly specific to TACC or basic Autopilot. I don't get any of that, I'm all set up in NOA / FSD, with all the options tweaked just the way I like them.

Oh, yeah, and the car took vengeance on her and directed her on some crazy half hour detour through the hills rather than the freeway, when she thought "Navigate Home" was in effect. We were both scratching our heads - it WAS a beautiful scenic drive - but that's when I saw her clock was showing 6:14 PM, and mine would say 18:14, and I realized what was wrong. Heck, the car won't even open for her unless I'm around, with my phone ;-) Come to think of it, maybe there's some sensible AI at work here. Don't forget that all the settings are specific to each driver profile!!!
.