-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Washington Post: How Elon Musk knocked Tesla’s ‘Full Self-Driving’ off course
- Thread starter FloridaJohn
- Start date
E90alex
Active Member
Everyone whining here about Musk being stupid removing radar, and they are so far behind on FSD and such, should watch this video.
I watch this YouTube channel regularly, and I watched this one earlier. While I generally agree with the sentiments and conclusions regarding Tesla being a leader and not a follower, I have to say that he leaves the strong and wrong impression that only Tesla, among the serious autonomy contenders, has a strongly NN-based approach. One is left with the impression that everyone else (outside of bit players like Wayve) is still on stuck on "Software 1.0" techniques.
Though I can't claim to be an Insider, I believe that most if not all of the serious Robotaxi companies are also heavily ML/NN based, with a lot of detailed differences regarding system architecture and partitioning - along with the well-known differences in the Tesla generalized driving approach versus the heavily mapped and geofenced "virtual rails" approach. I think this NN error is kind of an unfortunate flaw in otherwise reasonable commentary, and I hate to see him set up such an easy criticism of the video.
Yet it still somehow worked better than Vision only
It may have in some situations but not in others. I might be in the minority here, but when radar was disabled, my highway drives became PB-free. no more false positives at overpasses and signs.
I don't necessarily agree that Tesla Vision is a complete dead end; it's just that it wouldn't perform the way we want it to any time soon. Better to use some "crutches" along the way if economically feasible to get a decent product out the door, assuming radar can do that - that's actually still an unknown. Then someday if vision AI gets to the point where the other sensors aren't helping anymore, THEN get rid of them.
It's the obvious compromise in approach that Elon is stubborn about, but good to see the HD radar coming back.
Good point. Although the one that gets trained with a ton of real world data wins.
What is the purpose of a well crafted NN with no real world data? So in that sense nothing comes close to Teslas approach
I know radar isn't the only controversial thing with Tesla's design, but since we're talking about it...
Everyone here seems to be assuming that it has to be radar on 100% of the time or it isn't worth anything. I don't see it that way. I could easily see using primarily vision, and then "augmenting" with radar when it makes sense.
For example, the car is using vision, and it comes upon another car just ahead. The vision system now recognizes that there is a lead car. Then the radar is activated and is used to set an accurate follow distance. When the lead car pulls away or exits the road, the Tesla can ignore the radar returns. That would be a sensible type of "sensor fusion" to me.
This simple method would let the car get back to the '1' follow distance, and would greatly improve reaction to the lead car when traffic is heavy or when you are in stop-and-go traffic. I find the current vision implementation of following a lead car too "spongy" and it does not follow the lead car accurately or naturally. I mean, it works, but with the addition of a strong distance measurement it would be so much better.
Everyone here seems to be assuming that it has to be radar on 100% of the time or it isn't worth anything. I don't see it that way. I could easily see using primarily vision, and then "augmenting" with radar when it makes sense.
For example, the car is using vision, and it comes upon another car just ahead. The vision system now recognizes that there is a lead car. Then the radar is activated and is used to set an accurate follow distance. When the lead car pulls away or exits the road, the Tesla can ignore the radar returns. That would be a sensible type of "sensor fusion" to me.
This simple method would let the car get back to the '1' follow distance, and would greatly improve reaction to the lead car when traffic is heavy or when you are in stop-and-go traffic. I find the current vision implementation of following a lead car too "spongy" and it does not follow the lead car accurately or naturally. I mean, it works, but with the addition of a strong distance measurement it would be so much better.
stopcrazypp
Well-Known Member
Yep, others are also using NN in other logic, perhaps even moreso than Tesla. The gist I got from FSD Beta was Tesla actually was using NN primarily in perception and most logic was still hard coded, so they actually were behind in this regard. Only recently are they pushing to use NNs in a wider way and doing away with some of the code that were manually coded.
My understanding is that V11 removes the follow distance all together, giving you the three modes: Chill, Average, Assertive. So we can finally move off the "Follow Distance of 1" debate that's been raging for over a year.
Yes, in 2 years I never set my follow distance to anything but 7. Tailgating is for chumps. Anyway the media is predictably going after Elon Musk and all of his companies now, having made an about turn after he opened up the Twitter cabal and proved the collusion between government and the media including Twitter to shape opinion and control speech in violation of our intrinsic right to free speech.
COS Blue
Member
It blows my mind all the people defending the removal of radar, saying this was just a hit-piece, etc. Since the removal of radar, I cannot even use cruise control if the sun is in front of me, lower than the roofline. It just says it's lowering the speed due to "poor visibility", even though I can see just fine. When I had radar, this literally never happened. Due to this issue and the ongoing PB events that freak out my passengers, I often just have to turn off cruise control now. This is basic functionality that cars have had since the 1980s.
Tesla has literally removed functionality from my car in order to pursue some sort of purist "vision". And the rationale is incoherent. Cars don't need radar, they just need cameras because that's what humans have! We're adding a new better radar in our new cars! But we're taking it away from your car whether you want us to or not! Trust us, it'll be better! In the short-term, however, we are going to take away some capabilities! Our autopilot system is safer and more capable than a human! But please don't try to use it if you can see the sun in front of you!
The article confirmed exactly what I've been thinking for a long time, and also enlightened me on things I did not know. It's good journalism.
Tesla has literally removed functionality from my car in order to pursue some sort of purist "vision". And the rationale is incoherent. Cars don't need radar, they just need cameras because that's what humans have! We're adding a new better radar in our new cars! But we're taking it away from your car whether you want us to or not! Trust us, it'll be better! In the short-term, however, we are going to take away some capabilities! Our autopilot system is safer and more capable than a human! But please don't try to use it if you can see the sun in front of you!
The article confirmed exactly what I've been thinking for a long time, and also enlightened me on things I did not know. It's good journalism.
stopcrazypp
Well-Known Member
People aren't disputing removal of radar has made TACC/AP worse for some people (although of note it did make things better for some people too), but the article was not about that. The article was about removal of radar making FSD worse, when the opposite has happened: Vision-based FSD Beta is a ton better than the version that had the older radar unit still running.
+1
Especially the latest version seems to be a solid improvement. Even some of the past highly skeptical FSD beta users have changed their tune with this version.
COS Blue
Member
Well the article did not only deal with FSD, even though FSD is in the title.
Disabling radar affects autopilot (TACC) as well as FSD. If the strategy is to ruin Autopilot (TACC) because it makes FSD better (which I'm skeptical of), that's not a good strategy. I find it hard to believe that a low sun cripples cruise control, but is no problem at all for FSD.
solid improvements keep on coming but still nowhere near Level 3. something Elon has promised for years now. (never mind the L4 robotaxi talk).
Ya so?
This is not a solved problem by anyone so far. So Tesla is pushing into unknown territories and solving problems and pushing ahead with every release.
If you can’t see and acknowledge the progress they have made so far, and are just here to criticize their timelines and delays, that is fine.
Tesla and SpaceX makes something that is impossible to late. Musk said 2020, but if they get there only by 2025, mankind will survive and world won’t fall apart.
only difference... Tesla is also the only manufacturer who is selling and has been selling "FSD" for a pretty penny since 2019 (so *years* ago...) but up to now it still remains a glamorized Level 2 Beta test. Let me know when FSD is at least "hands-free" like GM SuperCruise/ Ford Blue or Mercedes-Benz in *certain* conditions... we aren't even there yet for FSD from a regulatory perspective.
mspohr
Well-Known Member
So, the fact that FSD keeps getting better is a significant action.
Similar threads
