Ditto.
[Looks over thread.] The last Tesla driver update was Feb. 10, so just over one month ago. Excluding the video I posted about, the only other news in the thread since then was an update on a Waymo getting a ticket for running a red light. My impression is users updating everyone on the marginal improvements to their Tels self-driving software is roughly 2/3 of recent posts.
There’s a lot more discussion around the “Tesla didn’t detect the wall” video we just watched, but I’m not knowledgeable enough to figure out what’s accurate or what’s meaningful. Still, the implication is that it was an inaccurate test.
I mean, here’s someone saying the debunking has been debunked.
Tesla fans exposes Tesla’s own shadiness in attempt to defend Autopilot crash | Electrek
There have been two or three minor updates since then that weren’t worth discussing so we didn’t.
Again, I am very open to discussing all of the progress that other manufacturers have made.
I haven’t wanted the video yet, though I have seen some of the back and forth. Regardless, though–even that article acknowledges that this is (at most) an Autopilot test, not FSD. They’re drastically different products at this point. I don’t think Autopilot’s gotten meaningful updates for a few years now.
I’ve no idea how FSD would behave here (though probably not that well since fake walls aren’t likely to be in the training set), but regardless, Rober’s video doesn’t have anything to do with FSD.
In general, vision systems are very good at inferring depth today with multiple video feeds. There’s no fundamental limitation here.
This guy did another test with FSD:
Short version: a Model Y with HW3+FSDv12.5 didn’t see the wall. A Cybertruck with HW4+FSDv13.2 did see the wall and stopped smoothly. The Cybertruck doesn’t have wildly different cameras or anything compared to a new Model 3/Y, so this probably comes down to the HW and FSD version difference.
So obviously LIDAR isn’t necessary, but at the same time handling these edge cases does seem to be still in flux. FSD on HW3 is up to 12.6, so possibly things have improved. Or maybe HW4 will be necessary.
His decision to bring in the Cybertruck came at the last moment, as the day was ending, and the lighting had changed. Many commenters thought, as I did, that the camera’s ability to detect the wall might have been affected by that, either because of the difference in the light levels around the wall, or by things like the light reflecting off of it from the CT’s headlights.
Several commenters suggested that he repeat the test with the CT at the same time of day as he had with the Model Y.
I was wondering why he didn’t let the car crash through the wall on the last run, until I caught an overhead drone shot that revealed that the wall wasn’t self-supporting, but was leaning against a large panel van parked behind it. 
I agree it’s possible that the shift in lighting made a difference, but from the shots made inside the car, it didn’t look that dramatic. May have even favored the CT. Hard to be sure.
I would like to see a better controlled test. This was better than Rober’s test, but could still stand some improvement.
And if we’re really testing edge cases here, trying to see if LIDAR has a fundamental advantage, we should be stressing things that LIDAR is potentially bad at. How about a polished mirror car? Disco balls in the road? Etc.
It’s not a good look for Rober that the CEO of Luminar (the LIDAR company in his video) is friends with Rober and donated a bunch of money to Rober’s charity. That’s not necessarily evidence of ill intent, but it makes it hard to trust Rober on this.
I also laughed when I saw the Home Depot truck behind the wall. I think they could have built something out of PVC or whatever that stood up on its own. I sure wouldn’t have conducted the experiment like that…
Another test by the same guy:
This time the lighting conditions were the same (time was within 10 min) and they used a Model Y for both tests, a 2024+HW4+v13.2.8, and a 2022+HW3+v12.6.4. Same result: HW3 doesn’t see it, HW4 does.
And for your satisfaction, he does crash through the wall on the last test 
(no panel van this time!).
So that is pretty interesting. Could suggest that HW3 is running into its natural limits, though it may have more to do with v12 vs. v13. Either way… no LIDAR needed on the latest version.
Incidentally, I did put on some more miles on my own car with v12.6.4, and continue to be impressed. Drives for hours at a time with zero intervention. Got stuck in pretty bad traffic, going ~30 mph average for like 4 hours, with only one notable intervention, which honestly was just the car being a little too aggressive than I’d like going around a car turning right. Would have been fine unless the car changed their mind about the turn. And that was early in the drive.
Lots of impressive little things, really. Once I noticed the car starting to hug the right of the lane. Before I had time to think about it, a group of lane-splitting motorcycles roared by on the left. The car had given them an extra 6-12 inches of space.
Probably half a dozen times I had this experience: I’m in the right lane of the freeway and someone is coming up from an on-ramp and will end up at the same point. One of us can speed up or slow down… or I can change lanes. The car changes lanes each time, going back to the original lane when it’s safe.
Passing cars in general felt very natural. It’s hesitant enough to feel safe, but when it does change lanes it’s fast and quickly comes up to the speed of the faster lane. Doesn’t cut people off or force people behind to brake.
I have v13.2.8 and I haven’t had an intervention in months. My biggest complaint was that when two lanes merge into one it was way too aggressive and didn’t take turns but that’s been solved. My now biggest complaint is that it doesn’t move out of the way when it’s going too slow in the left lane of the highway. I pay close attention (which I should be doing anyway) and have it move over by signalling right to avoid that.
My biggest source of interventions remains map issues. It attempts to miss turns, or gets in the wrong lane.
I still really don’t like the speed control on v12. I think I understand what it’s doing, though. In previous versions it was more like cruise control, with the set speed being the speed FSD would try to drive. It would slow down for traffic and curves, but when able, it would travel at the set speed.
Now there is no more set speed; it is a max speed. FSD will go at whatever speed it feels like, staying at or below the max speed. On my neighborhood roads, where it wants to go 15 over the speed limit, the max speed works as a set speed. On the highway, where it wants to go 5 below the speed limit, I’m constantly having to make it go faster.
Best I can figure, it’s been trained on the most annoying drivers possible. People who want to go 50, regardless of the speed limit. People who can’t hold their speed going up hills. People who vary their speed on randomly on the freeway.
I’ve read there are supposed to be rules for above and below 55 MPH speed limits, but I haven’t been able to get that to work. Perhaps that is a v13 feature.
So yeah, v12 seems really good, it just needs to get the speed stuff sorted out.
Yeah, the set speed thing is annoying on the highway. I’ll turn down the max speed sometimes because there is construction of something. For example the speed limit might be 65 but I’ll wind it down to 60. If a little later the speed limit changes to 55, it will reset the whole thing and have a high set speed again.
That was one major intervention on my trip but I honestly can’t blame the car. There was a very strange section of construction work where the highway separates into two sections, with one of them having access to the ramps. It’s obviously not on the map since it’s construction. The car just happened on that section due to being in a right lane, but it got a little confused navigationally and tried to take an off-ramp. I think it probably would have figured things out and gotten back on the highway, but I intervened instead.
Surprised at this one. There’s a somewhat steep grade on my trip which slows most cars, but FSD was passing people on the right at a solid 65.
I didn’t really have any speed issues at all. Basically did what I want–drive maybe 5 over when traffic is light and just go with the flow when it’s dense. It’s a little more hesitant than I’d like about getting unstuck behind a slow driver in the right lane, but I had it set to chill.
FSD is good at this. If the traffic is dense, but still traveling at full speed, then FSD does a great job keeping up. The speed limit is 65, traffic is going 70-75, FSD max speed is set to 75, and it handles it perfectly.
Then traffic thins out, and without making any changes to the max speed or speed limit, FSD decides to go 62.
The uphill slowing is really frustrating, because the car has more than enough power to climb at any speed, and the responsiveness is instant, so there’s no reason to slow down at all. That’s how it worked pre-v12. I could set 70, and the car would go 70 uphill, downhill, whatever.
I think it needs some kind of rules based system mixed with the AI system. It has to have something like that, because it makes full stops at stop signs, while I’m sure the training data rolls most of them.
Bwahaha, it drives like my dearly departed mother? No thanks.That kind of explains the Teslas that I see hanging out in the left lane going below the flow of traffic, though.
To be fair, I see a lot of basic human-driven ICE cars doing the same. Lots of people like to camp in the left lane regardless of speed.
IMO it feels safer to them to not have people whizzing by on the left. Because those passing cars are near the driver. Cars whizzing by on the right are 5 feet farther away. Soo much better! And of course timid drivers are usually slow drivers. So that leads to left lane camper-outers being especially slow on average.
IMO YMMV.
Only if she was the kind to drive 50 in a school zone, and then drive 50 in a 55 zone…
I’ll give a serious answer about FSD here. v12, and even v11 before it were good at moving to the right when a car got close enough behind in the left lane. What they are frequently not good at is moving over preemptively. So FSD moves to the left to pass a slower car, stays in the left lane because there’s no reason to move to the right (you know, except laws, common courtesy, and probably some religious doctrines), then a faster car comes up behind, perhaps it had just moved into the left lane to pass the slow car on the right. The faster car is forced to slow down for the Tesla, and then the Tesla moves to the right, letting the faster car pass. An attentive human driver would have moved over earlier, so the passing car didn’t need to slow down.
It can also get it wrong in the opposite way. FSD moves to the left to pass a slow car, which is actually several slow cars. There is a car behind the Tesla in the left lane, perhaps not even going faster, just close because of the density of traffic, FSD will then move (if safely able) to the right into the group of slower cars. Even a conscientious human driver would have finished passing the group of slow cars before moving over. And considering that FSD changed lanes in front of the left lane car in the first place, it was probably not going faster than FSD wanted to go.
So that is a really long way of saying that if a slow Tesla is camping in the left lane, it is probably for the same reasons a slow Subaru camps in the left lane (obliviousness, righteousness, and cannabis are the reasons in Colorado). FSD is bad at predicting traffic flow, and occasionally wants to drive in the left lane for no obvious reason, but is good at getting back to the right when someone wants to pass.
That is kind of how she drove in her later years. She had “her” speed, around 45-48mph, and she drove at it everywhere. If I rode with her on the freeway, I’d gently remind her that she was causing a traffic problem when she got out of the slow lane.
Well, the last part is an improvement over most of the folks hanging out in the left lane.
A pretty clear-eyed article on self-driving cargo trucks starting to actually roll out in Texas. I’ll be interested to see how this goes…I have hopes for this.
Having made the drive south on 35 many times, the number of traffic jams should keep speeds relatively low most of the trip.