The fact that you can't use it without performing the full OEDR yourself? I think I've said everything there is to say.
In a dark room, would you say it's more reliable to measure a distance with a laser tool, or a guess from a 2D-still image?
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Autonomous Car Progress
- Thread starter stopcrazypp
- Start date
-
- Tags
- Autonomous Vehicles
In a dark room, would you say it's more reliable to measure a distance with a laser tool, or a guess from a 2D-still image?
Why? Any example provided you will just brush it off to planning or NN.
You seem convinced that Beta is close (within a year) to being solved. Shouldn't the burden of proof be on you? FSD Beta has gotten better since I started testing on 10.3, but at a very slow rate. Even at a higher exponential rate that Beta has increased in those 2 years, it won't be "solved" in the next 2 years. Why are you convinced Tesla will figure it out in the next 12 months?
EVNow
Well-Known Member
OP is in Europe - so its all hearsay and YT. No actual experience.
willow_hiller
Well-Known Member
A useful mental exercise to separate the perception performance from the planning performance is to imagine driving your car with the FSD Beta visualization as your only window into the outside world. I like to imagine piping the visualization into a VR headset, and having it able to rotate around the ego as you turn your head from side to side.
I think I would be able to drive pretty decently with the visualization as my only view. I would definitely drive cautiously, and probably falsely brake more often that FSD Beta does, but I don't think it would necessarily be unsafe or impossible.
I think I would be able to drive pretty decently with the visualization as my only view. I would definitely drive cautiously, and probably falsely brake more often that FSD Beta does, but I don't think it would necessarily be unsafe or impossible.
powertoold
Active Member
Every time I disengage or see a disengagement on YT (I watch many many videos), I think about the rationale for the disengagement. Is it a lane uncertainty issue? Is it dangerous? Is it user preference? Is it a mapping issue (wrong speed, wrong lane, etc.)?
With 11.4.2, I rarely if ever see an issue that makes me believe that computer vision perception isn't there yet.
Right now, it doesn't seem like LIDAR is offering any additional benefit. The problems with fsdb are planning and mapping related. That's why I said the CV has won because a year or two ago, people thought it was impossible for CV to be good enough at distance and speed estimation. Just look at Chuck's turn, people thought it would be impossible for fsdb to perceive cross traffic at speed.
I don't think Lidar is a must, but it absolutely is more than just planning. The car gets confused with distance and speed estimation in complex situations. I assume that will get better, but that's why intersections, busy city streets, curves with stops signs, etc. present so many more issues.
There's a lot of issues on 11.4.2 where there's a distant island in the road after a curve and the car thinks it will hit it or the example I've posted twice, a stop sign with a car turning left, but is over the line. The car completely stops and the wheel stutters back and forth. You have to disengage, pressing the pedal makes the alert to take over.
It's making progress, but slowly. I think vision can be adequate, but I do question the camera position with many of my creeps, especially with an obstacle blocking the line of site. The car can't lean forward like we can, so it get's dangerously into the lane where I had to stop and reverse or had cars honk at me or swerve away.
Something that I and others have experienced on 11.4.2 is sometimes the car has confidence that it can make a UPL, but misjudges speeds and if you either don't stop it or floor it, there would be an accident. Someone posted an incident of FB where it did it, they let it happen and the car had to slam on brakes (100% their fault), but it's newer behavior where it has too much confidence and misjudges speeds.
Last edited:
I've recently been in a Waymo and I have been trying FSDb 11 as a part of my research and general interest. Hoping to try a Cruise next time in the US. And you my friend said in the other thread about robotaxi that HW3 might get it around 2030? Sounds to me we're in agreement with you being only slightly more optimistic than me in camp "never". You also said the Tesla will likely upgrade the hardware even though the company clearly has stated there will be no upgrades of older AP HW...
Now, which is it? L4 in 2030 or 2024?
Last edited:
EVNow
Well-Known Member
How ?
As to robotaxi I only chose 2030 because that was the closest, because as I said Never is a long time
BTW, when I bought FSD in 2019, I said robotaxi is unlikely in the short term, let us see where we are in 2030. I've not changed my opinion. You can search for that post
I still think it is likely Tesla will crack FSD (in some form) earlier than Waymo expands to top 100 cities of the world - irrespective of how many million LiDARs they put on their cars.
There are so many other complicated issues with Robotaxi beyond the car being able to drive without a driver, which I don't think will occur before 2030. I think by the end of 2024, we will have a confident lvl 2 ADAS where ****-ups are more rare and maybe some lvl3 type hands free on the interstate.How ?
As to robotaxi I only chose 2030 because that was the closest, because as I said Never is a long time
BTW, when I bought FSD in 2019, I said robotaxi is unlikely in the short term, let us see where we are in 2030. I've not changed my opinion. You can search for that post
EVNow
Well-Known Member
Perhaps HW4 on dry roads and otherwise a minimal ODD (daytime, no low sun, limited speeds perhaps etc). I think highway speeds are harder than you think as L3 requires a MRM in case of failed handover, and you can't just stop on the highway (it's illegal). There is a reason for Waymo doing 60 mph rides last.
I have a lot of friends in the US. Some have Teslas, a few have FSDb.
EVNow
Well-Known Member
But how often do you travel to US ... after all FSDb has been out only for 18 months or so.
Anyway - the point is, critiquing 11.4.2 without actually using it - and not mentioning that you have not experienced first hand - is unseemly.
I said some lvl3 hands free, which we have to an extent (Mine goes 2 min on the highway and maybe 30 seconds in cities without a nag). I don't mean go to sleep lvl3 or full control. I agree there are additional complexities for that. I know Elon said at one point he's not interested in limited level 3 like Mercedes is doing, but maybe other competitors will force Tesla to comply? I don't know.
The main difference between L2 and L3 and above is that the system performs the full OEDR in L3+. This means you do not have to participate in the driving task (and can watch a movie, work or read a book). In L4 you do not have to take over when it starts to rain for example, because the system won't leave the ODD rolling. L3 is typically a more limited ODD (no rain, no low sun).
I think there is a chance of L2 hands off with some lenience of eyes off, so better than what Ford/GM is offering, but not full level 3 where you can watch a movie, sleep, whatever.
Again, right now FSD B 11.4.2 will allow you to go 2 full min without hands if you look straight ahead and some leeway, compared to City Streets, to mess with the screen without hands (15-30 seconds). All level 2 hands-off in the US currently won't allow any flexibility to look away.
It's binary though. Either you're legally driving or you're not. I can "look away" briefly for a few seconds even on standard AP. That doesn't mean I should, if you read the manual. But here we are after years of Elon's autonowashing. L2 is still L2. When Tesla takes on manufacturer liability for the system, we have L3+.
But they've loosened the requirements in the car, specifically on the interstate with 11.4.X. Yes, you are legally driving, but the system is becoming less strict on how much you pay attention, depending on the situation.
This stuff matters. It's not black and white. There's a ton of people who feel Blue Cruise is superior just because it's hands free...and it won't even change lanes, yet. Loosening the requirements is a big step in end user satisfaction, but the system has to be able to handle it safely.
I'm saying that by the end of 2024, I could see FSD Beta, on the interstate, allow hands free with additional leniency to mess with the screen, look around, etc. I never said it would be level 3. I even said Elon said he had no interest in that.
Typically you would have stricter requirements on the DMS when you allow hands off (see Ford and GM for example). Unless you want to be charged with manslaughter, please watch the road (ie perform the OEDR) while using an L2.
Similar threads
- Article
