I guess they think their computer vision and AI can catch up to and exceed the visual processing capabilities of vertebrates with their millions of years of evolution. I think they are wrong. Multi-media sensor data fusion is the only sane path for the foreseeable future, in my lay opinion.
They are wrong. A truly unbelievable coincidence coming across this article, I just got rear-ended by a Model 3 this morning.
Low speed, I'd figure about 10 mph tops: A softball to end all softballs for AEB, and somehow I got hit with considerable force. We've had radar based systems that would have avoided this since the 2000s.
I work in the self-driving space and while it's not an apples to apples comparison because of cost, all I could think is how our sensor stack would have allowed recognizing the car I stopped for, and probably the car it stopped for, let alone my car stopping.
I've already harped on how stupid the "FSD Beta" is, but I quite literally the absurdity of it shoved in my face just this morning only to come and read this. Why is Tesla even being allowed to run this circus at this point?
My entire point is that even outside of FSD this is a scenario a base model Corolla has the technology to handle: AEB
You don't actively engage AEB, at the kinds of low speeds involved AEB would have engaged and at least helped stop the car.
This was traffic that was going maybe 20 mph, and the car still hit me solidly enough for my car to lurch forward and almost hit the one in front of me
-
But as an aside, this is why Tesla's "Full Self Driving" shouldn't be on the market: It is always human error.
Even if FSD was enabled (I hope it wasn't) it'd be the human's fault for not reacting.
> Teslas AEB works but if the driver has their foot on the accelerator it won’t.
Seriously?!? Tesla drivers learn to basically always have their foot on the accelerator, because the car slows quite aggressively with the accelerator all the way released.
IMO the one-foot model makes it quite hard to drive without lurching a bit and without unnecessary accidental light breaking while cruising. I often imagine that careful haptic feedback on the pedals could enable a much better one-pedal or blended braking experience.
I feel that may be irrelevant - my 2019 honda odyssey will stop me no matter what. All lane departure / following, cruise control, etc can be off - but automatic emergency braking / anti colission will work. I would not have it otherwise - I do not have any automated stuff in my old wrx, but I wanted all the safety always in the car with my family.
(it is crazy to me how many brands only have the good safety stuff as part of some ridiculous 9000$ sunroom and leather and Sirius xm package / trim ;-< )
It is completely irrelevant, but that's not stopping people from downvoting the anecdote, seemingly because they don't realize (or don't want to accept) the fact that Tesla has regressed so far that they're starting to miss the bar that earlier versions of the same vehicles set with AP1!
Contrary to some of the replies, at low speeds AEB does override the accelerator, unlike the highway ACC component of AP, so it should have engaged.
For the record, I just took delivery of a Model 3 and seemingly all of these safety features can be toggled off in settings. Because of how complex these systems are, it wouldn't surprise me if someone turned off or configured automatic emergency braking because they thought it would make autopilot better or something.
I CAN turn off AEB in my car, but I would NEVER do that, because it's only shown itself to be very reliable. It's radar based. I never get false warnings from stationary objects and I've only had a single warning that I would consider phantom.
The IIHS tests vehicles with automatic emergency braking at 12mph and at 25mph.
When these systems were new, many vehicles were only able to decrease collision speeds. These days, many vehicles completely avoid crashes at both 12 and 25mph.
I tested it a little bit during the 2019 Honda Odyssey test drive. It seemed to work.
FWIW, Then unintentionally I had a bit of real world confirmation: a few months ago a car in front of me in slow traffic made a very sudden and complete stop.
I stepped on the brakes; after a second, as I slowed down, the brake feel completely changed and became harder, and I realized that I hadn't been alone - initially, car was braking for/with me, as confirmed with the massive orange letters on the dash board. Then as the system felt it was out of most immediate danger, it released and it was just me braking (that was the brake pedal feel change).
I do occasionally have it engage stage 1 of 3 (1 Visual -> 2 Audio -> 3 Auto Brake). It happens once a month predictable on very slight curves when it thinks I may be accelerating into oncoming traffic.
>an catch up to and exceed the visual processing capabilities of vertebrates with their millions of years of evolution
I mean tons of technology outperforms humans by a mile... plenty of other computer vision systems already outperform people so not sure why Tesla would be exempt here due to 'evolution'.
We don't have any computer vision system that outperforms mammals, let alone humans, at general vision, Tesla or not - especially on real 3D vision. We do have computer systems that outperform humans at extremely specialized tasks, such as facial recognition in photos.
Again though - these sensors performed a highly specialized task - there is no inherent reason Tesla wouldn’t be able to outperform them with a specialized visual system
The problem of using computer vision to compute distances to moving 3D objects in any weather conditions is too general - it is precisely the generalized vision problem I was talking about.
Ultrasonic or radar or lidar sensors are much simpler solutions to the problem of computing distances, as they do not rely on computer vision.
Note that the reason this problem is very hard for computer vision is that there is simply not enough information in a 2D picture to get 3D distance information, even with parallax. Our eyes also can't solve this problem. Instead, our visual system uses numerous heuristics based on our inherent understanding of simple classical physics and the world around us.
For example, we recognize that a particular blob of color represents a car, and that cars are solid objects, where each part of the car moves at the same speed as each other part (this assumption breaks if something flies off the car). We recognize that objects throw shadows, and what effect shadows have on color, so we can very easily tell the contours of an object even if it's speckled in light and shadow. We also know what approximate sizes cars are, and that means we can immediately differentiate a far away car from a close one. We also know that cars sitting on the road are probably moving, while cars on the sidewalk are probably parked, and that again helps us estimate their movement based on relatively little data.
If you don't believe that depth perception is based on far more than parallax, just try to explain how come people/animals with a single eye can still estimate distances to a very good degree (not enough to be marksmen, but more than enough not to run into walls).
> there is simply not enough information in a 2D picture to get 3D distance information, even with parallax. Our eyes also can't solve this problem.
That is certainly false - perhaps with extremely limited amounts of parallax yes but that isn't how this setup is working. Humans certainly use heuristics not only because our brains are capable of it but also because a face only has a few mm of separation, with enough parallax you can get phenomenal 3d information out of a visual system. Replacing a single sensor here is certainly not too general for good performance but also my point again was never to say this was an optimal or perfect solution only to say evolutionary arguments are stupid here because we build machines and algorithms every day that operate better than evolution does.
Because driving isn't some narrow task. It has a lot of variability in conditions. Human brain is very adaptable, yet we require 16+ years of training before they can take the wheel.
I'm not disagreeing that the task is difficult but this topic isn't about general driving - this is about replacing ultrasonic sensors with visual systems. Evolutionary arguments are meaningless here since we outperform evolution all the time even with hand engineered visual systems, let alone ML derived ones.
Nah I think it's a logistics and cost thing. Adding several extra sensors dramatically increases the complexity of logistics, building the car and the overall costs.
