The idea that humans drive without RADAR so vehicles should be able to also is stupid. There are a lot of things humans end up hitting because humans can't see beyond the vehicle ahead of them. Automated driving should aim to be better than human drivers, not to just make exactly the same mistakes because it has exactly the same limitations.
-
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
AlanSubie4Life
Efficiency Obsessed Member
Right, but the issue is that with awesome radar and good cameras (and even lidar!) it will likely still suck. I think that is what the article missed.
We’ll see, though. No one knows how it is going to turn out, since no one has solved the problem.
A good radar would help of course. So seems like something to consider!
Very complicated; I have no idea what is going to work.
stopcrazypp
Well-Known Member
Well, to do L5 (or generalized driving), the car would need to be able to drive without radar even as good as humans do, so the idea was might as well focus directly on that than on what may be temporary clutch.
But obviously things are much tougher to accomplish that way and a mix of the two is still quite useful, especially for L2 (which is really where most of the observed regression is, in terms of phantom braking complaints, which is the only quantitative data the article pointed out).
Elon's claim (repeated by Karpathy in presentations) is the old radar units were not high res enough, so may make things worse for Vision in certain scenarios (basically syncing issues). The new radar supposedly will address these deficiencies. For the same reason, I don't expect radar to be reactivated on the old radar cars, I think the ship has sailed on that.
The only thing I really learned from this article is the claim that the radar removal was for cost saving, but notably didn't mention it as a parts shortage issue that might have halted vehicle production (unlike the USS which there is plenty of evidence, even for other OEMs, of removing them to avoid delays in production). Tesla seemed pretty alone in removing radar.
mspohr
Well-Known Member
Removing radar did save a little money but my understanding was that it took a lot of processing time to resolve differences between radar and cameras and the cameras were better than the radar so they went with the cameras.
lol. better how?
cameras at night or in fog can exactly see as well or as poorly as humans... radar isn't limited by night or fog...
That guy, Faiz Siddiqui, is a Elon Musk hating troll. His job is to write something negative about Tesla and Musk. He probably has authored around 50+ negative articles on Tesla and ZERO anything neutral or positive. There are 4 or 5 well known talking points these articles always regurgitate, with a biased view.
He keeps citing the removal of radar over and over again as a huge mistake by Elon. First off, lets give credit to the Tesla team that has accomplished so much what the industry "experts" over the years have said can never be done. If Elon is such an ego-maniac and treats high performing employees shabbily, then how do we explain that consistently Tesla and SpaceX and is the No 1 & 2 most sought after companies by engineering graduates?. How do we explain all the amazing technological leap from Software to manufacturing to drive trains to battery/cell technology they have achieved.? So lets not assume he is a fool and the reporters are geniuses.
Now we know the radar tech they had at that time was pretty much useless to advance FSD. The noise to signal ratio is just too much, and it just simply took too much CPU bandwidth to extract useful data and merge it with vision data. There is a new radar coming up that seems to have the potential of adding value. If Radar and ultra sonics are the panacea then where is the L5 FSD from other companies, that have been working on this far longer? Remember Musk and SpaceX engineers have enormous experience using Radar and Lidar in Dragon space capsule. They know what they are doing.
Everytime when Tesla/Musk do something different from established 'wisdom', there is so much of ridicule and hate thrown at Musk.
Two examples: Large number of Cylindrical cells when the industry wisdom was lower number of larger pouch cells. When Toyota was pushing fuel cells, Musk called them fool cells. He was right on both. I can give you 10 other examples where the 'experts' said he was wrong, but turned out Musk was spot on.
And then of course you can always get anonymous and ex-employees to reinforce any negative view point you want to throw at a company. Like the employee/Autopilot tester who was going overboard in his YouTube videos claiming FSD is a killer, and was rightly fired. I can always get a disgruntled employee to say 'Musk drinks babies blood for breakfast'.
Autopilot employees keystrokes are monitored - OOhhh evil Musk is a control freak !! Turns out, they are measuring the effectiveness of manual labeling to see how to improve that process, especially when compared to AI based auto-labeling. It is not to measure their time at the desk. It is measuring the start times and end times of doing each label averaged over thousands of employees. You cant improve what you can't measure. !
There is absolutely nothing that is new in that article that we don't know. He strings a list of unrelated and irrelevant items, dramatizes them giving a huge negative color, to paint a monster picture of Elon. He is a presstitute. Look at all his previous articles you will get the picture. When Tesla was struggling to scale up Model 3 production, he wrote a lot of hit pieces.
How many of you remember Martin Tripp who was fired by Tesla for sabotage and leaking information to reporters? The entire media stood behind him and ganged up against Musk and painted a villain picture. Turns out that guy was a real nutcase. This Siddiqui character was wallowing on that incident and wrote many hit pieces. All turned out to be nonsense.
This whole article is not just a hit piece, but it is also a $hit piece.
He keeps citing the removal of radar over and over again as a huge mistake by Elon. First off, lets give credit to the Tesla team that has accomplished so much what the industry "experts" over the years have said can never be done. If Elon is such an ego-maniac and treats high performing employees shabbily, then how do we explain that consistently Tesla and SpaceX and is the No 1 & 2 most sought after companies by engineering graduates?. How do we explain all the amazing technological leap from Software to manufacturing to drive trains to battery/cell technology they have achieved.? So lets not assume he is a fool and the reporters are geniuses.
Now we know the radar tech they had at that time was pretty much useless to advance FSD. The noise to signal ratio is just too much, and it just simply took too much CPU bandwidth to extract useful data and merge it with vision data. There is a new radar coming up that seems to have the potential of adding value. If Radar and ultra sonics are the panacea then where is the L5 FSD from other companies, that have been working on this far longer? Remember Musk and SpaceX engineers have enormous experience using Radar and Lidar in Dragon space capsule. They know what they are doing.
Everytime when Tesla/Musk do something different from established 'wisdom', there is so much of ridicule and hate thrown at Musk.
Two examples: Large number of Cylindrical cells when the industry wisdom was lower number of larger pouch cells. When Toyota was pushing fuel cells, Musk called them fool cells. He was right on both. I can give you 10 other examples where the 'experts' said he was wrong, but turned out Musk was spot on.
And then of course you can always get anonymous and ex-employees to reinforce any negative view point you want to throw at a company. Like the employee/Autopilot tester who was going overboard in his YouTube videos claiming FSD is a killer, and was rightly fired. I can always get a disgruntled employee to say 'Musk drinks babies blood for breakfast'.
Autopilot employees keystrokes are monitored - OOhhh evil Musk is a control freak !! Turns out, they are measuring the effectiveness of manual labeling to see how to improve that process, especially when compared to AI based auto-labeling. It is not to measure their time at the desk. It is measuring the start times and end times of doing each label averaged over thousands of employees. You cant improve what you can't measure. !
There is absolutely nothing that is new in that article that we don't know. He strings a list of unrelated and irrelevant items, dramatizes them giving a huge negative color, to paint a monster picture of Elon. He is a presstitute. Look at all his previous articles you will get the picture. When Tesla was struggling to scale up Model 3 production, he wrote a lot of hit pieces.
How many of you remember Martin Tripp who was fired by Tesla for sabotage and leaking information to reporters? The entire media stood behind him and ganged up against Musk and painted a villain picture. Turns out that guy was a real nutcase. This Siddiqui character was wallowing on that incident and wrote many hit pieces. All turned out to be nonsense.
This whole article is not just a hit piece, but it is also a $hit piece.
Last edited:
But if camera cannot see, the worlds best radar is of no use. Because the primary set of data is from the cameras, and radar/lidars only provides supplemental data. You CANNOT drive even ten feet when the cameras cannot see, even with the worlds best radar.
So the notion that somehow radar helps when it is foggy is just plain wrong. In the end Radar (that was available then) only adds to noise and bandwidth, and that compute bandwidth is better utilized to process more photons from more precision cameras.
EVNow
Well-Known Member
Yes - I think the is the biggest thing all the haters miss. It is not an easy problem - infact it hasn't been solved.
Its like slamming Elon for delay in getting Starship working. Duh.
ps : We should also not underestimate the "hit job" nature of things. There are "journalists" whose only job, it seems, is to write negative things about Elon/Tesla. When they investigate and talk to former employees, for eg., how do we know they are not writing only negative things and deliberately censoring all the positives ?
stopcrazypp
Well-Known Member
Others already pointed it out, but for purposes of anything beyond TACC, Vision is a minimal prerequisite for driving, given our road infrastructure was developed for vision (things like lane lines, traffic signals, etc). That means the car can't safely drive in conditions it can't see properly (even if there are non-visual sensors). So that night/fog advantage isn't really necessarily the most relevant.
As such, the article doesn't point to that at all as a factor (rather it uses phantom braking as an example for radar superiority).
The main advantage that is most relevant (even in perfect weather and applies also to phantom braking) is radar does a great job of determining the velocity of moving objects. That's kind of the thing Vision had previously done a much poorer job of. They have since improved Vision a lot in this regard.
So I used to work on fast ferries and my employer was very proud of the fact that the ships could berth in zero viability (i.e. thick fog). We trained for this regularly, where the Master or First officer would berth the ship using only the radar display and not able to look out the windows.
So it is possible, BUT....
A ship berths very slowly in real time compared to a car. The last 50 feet may be at less than a foot second as the ship comes alongside (even in good viability). Clearly, this would not be acceptable in a car....
So it is possible, BUT....
A ship berths very slowly in real time compared to a car. The last 50 feet may be at less than a foot second as the ship comes alongside (even in good viability). Clearly, this would not be acceptable in a car....
E90alex
Active Member
Let’s not get ahead of ourselves be concerned about a hypothetical future with Level 5 autonomy and live in present reality where the system is meant to be Level 2 drivers assistance for enhanced safety and driver comfort.
Radar would still provide useful information and allow at least partial system operation in present times, specifically TACC and AEB, even if the cameras cannot see the lane lines for full Autopilot use.
With vision, it would not be able to see any better than a human in dense fog or other situations with reduced visibility. Which is how accidents happen in such situations, drivers can’t see far enough ahead to react to the situation. Radar can still sense cars or obstacles ahead even in zero visibility and therefore provide enhanced safety.
If radar was truly unnecessary then Tesla would not be backtracking and adding it in HW4.
Let's all demand to see Faiz's personal stock portfolio, and a complete list of transactions from him over the last 5 years. Does he own Tesla stock? If so, does he short sell after publication hoping the stock will fall in value?
mspohr
Well-Known Member
The problem with radar was that it had lousy resolution (i.e. analogous to a very low pixel count camera 16x16).
The new radars are supposedly much better.
E90alex
Active Member
The lousy resolution still allows for better speed differential detection vs the lead car than camera vision.
In any case, the main point is that Vision only is a dead end and not sufficient for the capabilities Tesla wants to achieve in the future. Actually, Vision only operation is currently still not up to parity with previously enabled low resolution radar nearly 2 years after the removal of radar.
mspohr
Well-Known Member
Such low resolution that you don't know what object you're measuring.
jebinc
Well-Known Member
That was just “Charlatan spin” to camouflage his true unspoken reasoning - cost cutting and implementation challenges. How about the Plaid+ line, “no one wants it as 400 mi range is more than enough”. Most of his sheep followed along and felt “well fed” from all of these lines of BS. And there are many… wait for the Cybertruck back-peddling “spin” when final specs and features (and price) are announced…
@uthatcher
@WilliamG
… another line of BS, please….
E90alex
Active Member
Yet it still somehow worked better than Vision only
I don't think people realize how bad classic automotive radar systems are, like a giant city bus and a can of Coke look identical from the radar's point of view. You have to filter out everything that's not moving at high speeds so you'd just run into stopped fire trucks.
I have a radar car with FSD. Before they disabled radar, it frequently got snow on it (I live in Michigan) and my TACC and other features would shut down.
stopcrazypp
Well-Known Member
Not for FSD Beta purposes. Most of the issues didn't seem to be perception related necessarily, and current Vision Beta seems to be a ton better than non-Vision Beta ever was.Yet it still somehow worked better than Vision only
For TACC and AP, there is anecdotal evidence that radar may have been better for phantom braking in some cases, but on the flip side, phantom braking has existed with radar cars for an extremely long time. In fact, the earliest AP phantom braking was directly related to radar for things like overpasses and large signs, which Tesla had to create a whitelist for.
Basically the article mixes these things in all into one, when there is more nuance than just saying having radar is superior.
Similar threads
