2 Problems FSD V12 Cannot Escape
ฝัง
- เผยแพร่เมื่อ 21 มี.ค. 2024
- Enjoying this clip? Watch the full video here:
• Tesla's MIND BLOWING S...
If you'd like to support the channel, you can find me on Patreon here:
patreon.com/HansCNelson
Shopping at Tesla? You can support the channel & save by using my Referral code:
ts.la/hans71453
PANELIST INFO:
Farzad Mesbahi
X Profile - x.com/farzyness
TH-cam - @farzyness
Chuck Cook
X Profile - x.com/chazman
TH-cam - / chuckcook
CHANNEL LINKS:
Hans' X profile - x.com/HansCNelson
Hans' TH-cam channel - @HansCNelson
Hans' Patreon - / hanscnelson
MY BUDGET SETUP:
Laptop: M1 Macbook Air
Camera: iPhone SE (Continuity Camera)
Mic: Shure MV7
Audio/Video Interface: OBS
Streaming Platform: Streamyard
Want to create live streams like this? Check out StreamYard: streamyard.com/pal/d/60406730...
#Tesla #elon #musk #elonmusk #teslaengineering #teslaenergy #teslaai #teslafsd #fsd #fsdbeta #autonomousvehicles #autonomy #neuralnetworks #machinelearning #engineering #deepmind
Tesla, Elon, Musk, Elon Musk, Tesla Engineering, Tesla AI, Tesla FSD, FSD, AI, AGI, Technology, Software, Computing, machine learnings, AI models, Engineering, ai, neural networks, machine learning, engineering, deepmind, fsd beta, autonomous vehicles, autonomy - บันเทิง
![Supernova - aespa エスパ 에스파 [Music Bank] | KBS WORLD TV 240517](http://i.ytimg.com/vi/bepqfR-pxY0/mqdefault.jpg)
Chuck Cook is arguably the best FSD tester to comment on the creeping problem, which would be made much better by having r/l cameras as far forward as possible, giving the car the best view of oncoming traffic. Nature has shown us this by evolving prey animals, such as horses and sheep, with eyes optimally positioned to see to the side, and just behind their noses, to detect possible oncoming predators.
Evolve the cars to have side/frony looking cameras up by the headlights up by the headlights, just like nature has taught, ASAP!
Doing that would open Tesla up to the expense of either retrofitting additional hardware to the cars or refunding those people who have bought the promise of full FSD capability. The third option would be to try and fight the inevitable class action lawsuit which would almost certainly happen. You can see why they are loathe to make a move on this.
I think Chuck might be wrong about the camera on the B pillar because the vision from the B pillar camera is perhaps 6” to 8” outward from the position of the driver’s left eye. This difference provides an entirely different angle and perspective than if the driver was in the same position in the back seat with his head duck taped to the headrest.
The Tesla engineers aren’t stupid. They’ve had these cameras in these positions for many years and had the opportunity to move these cameras or add new cameras for the past five or six years. I think Chuck is underestimating what the B pillar cameras can really see when approaching a high speed intersection.
Yes, it will actually make it worse.
Try driving with your head held to the headrest with an elastic wrap to simulate the displaced location of the B Pilar camera from the nose of the car. Now imagine the driving advantage if you could place your head on the front hood Tesla logo. Tesla could (and in the future will likely) do this with FSD. If I were you, I would listen closer to what the professional pilot is telling you. He is driving in 4D. The Air Force has had in the air combat FSD for years.
@@darwinboor1300 Yes, but we are talking about pulling out at an intersection where the cars are coming from the left. Someone in the back seat driver’s side with his head halfway out the window, will be able to see much more than you can imagine when sitting in the driver’s seat. This viewpoint along with the other cameras might be enough to get the job done.
Superb analysis guys! My thoughts exactly. I certainly couldn't drive my MY in the UK without leaning forward at most junctions!
Much of Europe.
I agree that even though there are 44,000 traffic fatalities in the US each year, if a nationwide Robotaxi fleet has one fatal accident every day or 365 fatalities each year, the media, politicians, legacy automakers, big oil, the the psychology of the American public will have a disproportionate reaction to each fatality even though the Robotaxi system would reduce fatalities by 99.17% (assuming the Robotaxis service replaces nearly all human drivers) and save over 40,000 fatalities each year!
Not to mention the lame stream fake media always fabricating FUD against Musk.
Yes 100% agree. That will be the key issue in the future.
Big oil will become more FSD friendly when humanoid bots will drive gasoline cars all by themselves.
The regulators probably wouldn't approve it, and even if they did, a huge percentage of society would irrationally be fearful of riding in a robotaxi. People are really, really bad at evaluating risk. The last three years clearly demonstrated that. There would also be a cost to Tesla. The government values the average human life at about $10M. So the 365 wrongful death lawsuits would cost Tesla $3.65B.
@@mb90821😊
why aren’t the cameras as far forward as possible, for example on the fender?
Thanks gentlemen. Chuck I agree with your points on the limitations of HW3 and V12. FSD clearly needs better side view imaging. Most accidents are not head on. V12 still needs a supervisor AI to deal with edge cases. This need is probably greater with end-to-end (E2E) AI. Eliminating the hard coded control planner eliminated many hard and fast safety limits. With E2EAI, using the current V12 architecture, those rules have to be learned for every possible variation of every edge case and they have to be learned without creating new errors including new errors in previously "solved" skills.
I'll supervise FSD any day for the added safety. For long drives I don't want to be without it. To get out of the car after an 8 hour plus drive and not feel fatigued is worth the cost of FSD. To make the drive with the added safety is invaluable.
Having said all of this, a number of major issues still need to be fixed. For safety issues, FSD needs to be able to pass a behind the wheel driving test, including school zones, traffic signs, alterred traffic patterns, unsupervised parking, and backing up to name a few. Until that time, talking about FULL autonomy is wasting time.
PS I suspect that the majority of accidents involving Teslas using FSD are the vehicle getting rear-ended. The solution to those accidents is to have all cars using FSD.
As someone who had a Tesla with FSD for a year and now subscribes my present Tesla, at least version 11 is nowhere close to eliminating the driver. But it is a very useful tool and I use it all the time. Waiting for version 12. I also believe that if Tesla lowered the subscription to $99 they would make a lot more money than the current price of $199. I believe a lot more people would be willing to pay the $99 even if the car still requires you to be in the driver's seat.
The pillar cameras have a wide angle and can see quite a bit tho. Same as you if you turn your head 90 degrees which almost no one does while driving.
But it still can not look around corners. That is why you lean forward.
Paralax is a huge interesting point. However, I believe that simultaneous multiple cameras and neural networks indirectly get to an equivalent valuable information.
Ai should be able to do it with one lens, especially if the object is moving as with cars if it processes consecutive frames. See this article as an example "Seeing depth through a single lens
Mathematical image processing creates a 3D movie of any scene, using just two frames from a stationary camera or microscope"
There is a reason as to why Waymo have gotten this to work:
They have cameras and lidars and radars and microphones like everywhere.
You should probably qualify that. Waymo loses something like $20 on every ride because their cars cost $200k, they operate a surveillance network on their small areas of operations and they employ a team of back-up drivers.
If Tesla is going to need more cameras, they could probably retrofit a small cluster on the nose looking forward, left and right, possibly with a single sensor and a special lens. They will figure this out.
@@labradore99 Well, Tesla really have nothing even close to what Waymo have so, you can not compare them.
Tesla of course needs more cameras, better placements, washing and wiping of the cameras, microphone, radar, lidar and more. They can not rebuild the car for this. Everyone who have a Tesla today are screwed. They will never have a self driving car.
@@Xanthopteryx Thanks, random internet guy. I'm sure you're right!
@@labradore99 Of course i am right here. Elon is just full of shit. I have been right for the last ten+ years and he has been wrong for the last ten+ years.
This apparent issue needa to be addressed by the Tesla team and Elon because I agree with Chuck about the required creep maneuver.
Level 3 will never allow you to sleep.
Level 3 means that you have to be ready whenever the car tells you too. That means that you will have to be awake and you have some seconds to take over but it is not like half a minute or so.
To answer Chuck, yes i think its safe to creep up to the white line. Drivers are made keenly aware of your presence as you creep closer to their driving path.
Of course standing on foot is dangerous, youre harder to see
I am curious about how dynamic the system is. Will each car Cybertruck, Semi, M2 require their own from scratch AI models with all new data because of computer, camera amount/position, size, steering changes. Or will it all be in same model and AI learns to drive according to the car and still retains all edge cases learned from past data.
Why not have a rod with a camera that could protrude out for a better view and then retract after the turn.
Excellent, very necessary discussion, when the human and the robot are taking half the responsibility, it doesn't account for others on the road, so it's a 4! Four factorial set of complex combinations of attention to driving one only vehicle. Surpervised FSD is the fact.
I see many challenges for FSD. One thing on liability on accidents could in fact play out similarly to the small aircraft industry. Example they stop making small craft in USA due to getting sued. So they solved it by making the aircraft kits. Then the assembler takes the liability usually the owner. So the future could be you install the chip and that one step removes liability for accidents. Or reduces it to enable manufacturers to continue to develop the tech. The other thing is establish a level of acceptable error similar to how we accept human error. That’s harder. And then new laws for j walking that could slow traffic or even protesters shutting roads down by walking onto the highway knowing cars will stop.
Everyone has forgotten that HW4 repeater cam went wide angle. I believe this fixed Chuck’s blind spot, assuming HW3 retrofit. The repeater is 1.3m forward of the b-pillar.
It does not since it does not matter how wide angle you have when there is an object in the line of sight. That is why you lean forward.
So what about a cheap retrofit, to add two extra cameras?
At some point the computing HW will be so cheap that it would feasible to upgrade the HW3 to be 10x more powerful, for basically free for the owner.
And we are very rapidly approaching that point.
Tesla can add two dash mounted cameras for the unprotected turns. This would also protect the lens from the elements and be more forward than a human head would be able to reach. It also can retrofit older Teslas this way.
Retrofit will never happen. Cables and contacts and the motherboard on the computer. It will cost way too much.
And also retrofit HW3 to a newer version will not happen either.
Everyone with a Tesla out there that bought the FSD software have been scammed.
@@Xanthopteryx The only solution may be my previous suggestion of slightly turning while creeping so the forward wide angle camera can see without going too forward. This would be the equivalent of a human walking forward and turning his head but the car will have to do it by going forward at an angle.
@@lym3204 It will not help.
Why?
Just make a drawing of the car with the camera angles, cut it out and then move it around on a drawn road with intersections and objects that are in the way.
And the forward wide angle camera is not that wide angle. First you can check your window and already there you see the limitation of view. It is only 60 degrees on each side, in theory. Remember that the image is very distorted at the widest angles.
And if you are on a one way narrow road with houses om each side, you can not even turn.
It is just a dead end they have, that they must improve to go further on, but they can't improve because that will mean then acknowledge that ALL THE CARS they have sold are unfit for the task and thus "unusable" for people, that was once promised that their car had everything to make it self driven.
That was of course a scam, and they have changed cameras two times since then, and computer have been changed what is it?
Three or four time now, or?
And radar is replaced with a new radar.
And new Teslas have a front bumper camera.
And they will change computer again. And probably add SODAR again.
And then radars in the rear will most certainly come. And washers for cameras. And wipers for cameras. And LIDAR. And microphone (directional stereo).
@@Xanthopteryxyeah, Tesla were too eager to cash in on the FOMO out there that they sold HW that was just not enough
Afternoon guys
Very interesting analysis of the current state of Tesla automation. I have only driven a tesla once from Hertz and was impressed with my experience. But I am surprised and rather shocked that the Tesla cameras are on the mirrors. I assumed they would be integrated with the headlights to give the car more forward vision. One of the things I loved about my Jeep was the cross traffic warning when backing. It seems clear that giving Tesla cars extended capabilities to go beyond the limits of human senses would be a step forward. Perhaps a retrofit may be possible. Creeping in a blind corner situation is certainly a seriously dangerous frequently encountered edge case. Thank you for your insights Chuck, and thank you Hans and Farzad for your work to help make Tesla tech issues understandable.
_But I am surprised and rather shocked that the Tesla cameras are on the mirrors._
Where'd you get that idea? That's completely false.
I thought that was what was claimed in the video as the reason it did not have a better perspective at blind corners. As I said, I do not know about the details of the designs having only rented a Model 3 for 2 days. I hope they do have cameras placed forward.@@Big_Ben_from_La_Mesa
@@daveoatway6126b pillar not mirror
Thank you for the clarification. I thought heard "mirror" when I listened to the excellent video.@@APatchworkCanvas
Two options.
We retrain drivers with camera views near the front of the car or move the driver forward and coincide with the camera view.
Either will work but I don’t think the goal is to have ai seeing different than a human because then correcting errors is too much of a guess. Can’t have those variables if ai makes a move because it can see better. We want trust and verify not blind trust at this time. Down the road a decade it may be different.
Computer vision cameras can determine depths in real time with a single lens. Plenoptic systems can determine depths within a single image frame. This is in addition to full spectrum sensor photon detection beyond the human visible range of light (e.g., infrared, ultraviolet).
The problem with single lens depth perception is that it is kind of very easy to fool.
@@Xanthopteryx Every perception system has its limitations. I still can't figure out something of the stuff Criss Angel or David Blaine do. Maybe Tesla needs to hire some magicians.
@@IndigenousEarthling101 Criss angel and David Blane: That is not because of your eyes but your brain (and also trickery and lies).
I agree with some of what you say. I don’t see why the car isn’t MORE visual than a human, more cameras, infra red (good for fog for instance), however there is one fallacy being discussed. He talks of pixels and range but forgets that human “pixels” are significantly less than on hardware 3 and 4. What enables us to see detail is a concept known as persistence of vision. We “fill in” the gaps by continuously looking. Moving images are filled in even quicker. Our eyes also have blockages that mean to see certain kinds of detail we must use the sides of our eyes. Tesla cameras don’t have this issue.
I personally agree with Elon that we don’t need sensors everywhere (after all humans don’t) and eyes are enough, but don’t we want our cars to be BETTER than us ? I want my car to see in the dark and and see further than me. I want it to calculate velocity and collision vectors (a Tesla already does this way better than we do ) . I think there should be a camera watching the road down low because it can adjust the comfort against the texture of the road. I also think that any “eye system” that is used for driving should be capable of calculating velocity and distance. I don’t know if you always need 2 eyes for this but humans certainly do. If cars do too, then there needs to be 2 eyes on each front corner looking at the sides, and 2 eyes at the back as well as the 2 eyes looking forward. If they can see in the dark that’s even better. This is not overkill, it’s just making the sure the car can drive “better” than us !
Great points. Reach out to Tesla or upvote on end of quarter call.
Sadly, cook was some urging for LiDAR and Radar. NOT an engineer or AI developer.
Maybe a little less doing left turns and going back to night school to stay current.
FSD Beta is not able to reverse yet. So also no 'reverse creep'
Unfortunately, got my first rear tire blowout, within first 5 mins in V12.3, when it negotiated a simple unprotected sharp left at 4 way traffic signal, in broad daylight, and malfunctioned(since the front tires made it fine) and ran over median curb. I can't afford to shell out $400 frequently, if this behavior persists😢. There was no imminent emergency situation, nothing unusual, a normal human driver cannot take care of. Also I was hugging my steering tightly, with eyes glued on the road infront and around me!. Unfortunately, I can't look down and take action in the middle of a turn, that FSD negotiated. I disengaged FSD promptly, reported incident to Tesla! and also filed a formal complaint to NHTSA! Hope the issue gets fixed in the next off the air update🤞😊❤️👍
Which model to you drive?
@@HansCNelson 2019 M3 AWD
Is Chuck forgetting that there are cameras in the windshield, too? Chuck, you aren't a Tesla engineer, I'm pretty sure they have a handle on this critical design detail.
The cameras in the front can not see to the sides.
Tesla have no clue at all. They have painted them into a corner.
IF they adapt the cars to actually be able to work then ALL the people out there with Teslas and the FSD software will go bananas and a class action is coming as a letter in the postbox.
Speeding is the main problem. We need traffic cameras everywhere to control this with steep fines. This actually works in countries that have implemented this,\.
I almost had two accidents this week. I ran red light and someone else ran red light.
Each almost accident makes you a better driver.
For purposes of pure "L5 robotaxi" purposes, the forward "peak-around" cameras are deemed necessary or at least a vast improvement to overcome limitations, they will not be difficult or expensive(compared to money-making value of robotaxi) to add and to adapt/train the AI to use them. The only real question is how "smart" can they make the HW3 computer, once the end-to-end NN is fully optimized. An alternative to adding cameras would be adding mirror(s) to the front of the car that FSD can pop up (and rotate if needed) when stopped at a blind corner. Training it to use these mirrors when needed should be no more difficult than training it on when to fold the regular rear view mirrors.
Another thing Chuck does not consider is how quickly the car will be able to reverse a "creep" the moment it spots danger. Realistically it might on rare occasions need to reverse suddenly pulling at most about 0.2g acceleration. Not ideal, not comfortable.
When it comes to turns with limited view, it may be better for AI training that the car has a worse view than the monitoring L2 driver. If it had a better view, could the driver trust it when they can't possibly see what the cars sees?
The solution for unprotected left turns, with high speed cross traffic is, is better road design. In my 20 years of driving in the EU, I still have not encountered one.
The problem does not only lie in that corner. There are tons of other situations out there that the car can not handle because of the sensor placement.
I believe it would be advantageous to have a JARVIS type system that could interact with the driver and give different levels of information. It could be as interactive as you wish or silent, but something besides text should be there to alert to certain circumstances/situations. The personalizations possible here are endless; from voice styles to ordering online.
I think only Elon could pull this off because he truly lives and believes in a sci-fi future.
C-V2X i.e connected cars technology is the super power senor needed for fog, heavy rain, blind corners, big semi truck obscured, etc. No other sensor can do what C-V2X can do.
C-V2X can prevent some disengagements that no other sensor can. The only limitation is lack of adoption, but cellphones can fix that if the FCC mandates it. Is Tesla getting ready for this critical new step? Mandatory FSD C-V2X AEB to end traffic accidents and fatalities
Maybe they plan on on outfitting all Teslas with cameras on the headlights?
Not retrofit and most probably not the new cars either. If so, they would have done it long ago.
why Better than human is not good enough?
So youre saying a Teslabot driving robotaxis would solve the creeping problem because it can lean forward with its camera.
The car needs to see the world like prey, not like a predator. A fish, not an owl.
Even if the roads are laid out properly, the objects that are blocking the view needs to be removed. The large SUV that parks on a corner will prevent humans from seeing traffic.
Why is this a problem you "cannot escape"? This could be fixed and you could even figure out how much it would costs. There are 8 cameras they could add two more and would if they're needed. Its possible internal to Tesla they don't think its necessary.
I think you are underestimating wide angle lens. Human eyes are not wide enough. Also pixel shifting allows perceptions similar to human eyes. Also I never put out my head to gain more visibility. Simply not necessary
Wide angle has no effect when an object is in the way. This is why you move your head.
And remember that your eyes moves CONSTANTLY and scans a very large area and with super high resolution. Hard to to a direct comparisation but roughly 500 megapixel.
FSD can never be completely safe if humans are still driving.
It will never be adopted first in the US as it is a very litigious society. Some small jurisdiction somewhere will be first once it is demonstrably safer than the average excellent human driver. Then the world will see a place with very low levels do accidents, where the vast majority or close to all of those that involve a FSD vehicle are clearly human error. They will have low cost taxi services and less congestion. Then it will begin to be licensed everywhere.
why don't they just put camers with parallax forward. duh! am I smarter than tesla engineers???
Overrated topic. Maybe it would be some improvement to reposition the left and right camera (with new drawbacks). But, never forget, even Hardware 3 is capable of 2000 processes a second. With enough training, the decisionmaking will be much faster than human, watching left and right the same time and finding it's gap.
Cook is wrong and the side cameras do not have to be mounted more forward and the solution can be achieved by FSD with less forward creep than a human driver. All FSD has to learn is to gradually turn left as it creeps forward then the wide-angle forward facing cameras will be able to see more than a human driver.
Agree with your view. Does Cook really think that Tesla hasn’t thought through the placement and type of cameras?
I agree. From the documentation it shows 360 degree view, no gaps. He is obsessed with the B pillar but the documentation shows overlapping views from the fish eye front cameras (there are 3 front cameras). But lets just play devil’s advocate and assume there is a blind spot. Tesla will just limit the areas u can do FSD, which is probably a tiny fraction of 1% of areas. And for people who want to go in those areas, you can have a bluetooth camera that sticks to your dash, window, or windshield. Or even an external mount camera add on. Come on this isn’t rocket science!
Reality does not work like that you know.
@@denero66 Of course Tesla have not thought of this. It is very clear. And the problem is that they can not really change now because that will make a lot of people angry and a class action is going to come straight away.
They have painted them into a corner.
@@Xanthopteryx read my comment, FSD rocket scientist.
One problem that FSD can not escape is someone else plowing into the car with FSD. The narrative today would be total failure of FSD for this situation.
Cook is losing credibility here. How about FSD can process 10-100x faster than a human. Duh
It really can not.
Heh heh, bye bye fsd.
Very impressive that an army of the worlds top engineers missed something obvious to a 4 year old
Hans, I see that this is becoming a trend, but that text that appears at the bottom of the screen is distracting and doesn't add anything to the video... it detracts from it. I simply don't watch videos that have this going on for more than a few seconds (at the beginning, to promote the upcoming video).
What I'd love to know from the FSD engineers is why, with v12.3, the steering wheel can make a 90 degree turn turning as smoothly as you or I would turn it, and with other 90 degree turns, the wheel operates like v8; very jerky and un-smooth; studdering?