Is it not? Yes we have 2 eyes, but if you watch TV you can process depth from various cues. Whereas if you're driving in fog and all you can see is headlights ahead, it's difficult to judge distance because you don't have those cues.
I suspect 2d only scanners is a symptom of a deeper problem. It's essentially (premature) optimision for an unsolved problem. And they're underestimating the problem. How long is it before a car can read the cues of a pedestrian that may or may not be about to cross the road without looking, or a dog or a teenager. Has that car stopped because there's something blocking the way, or have they just pulled over without indicating? Is the person just ahead that is driving unusually about to pull into your lane? The long tail of edge cases is just impossibly long.
My understanding is anything more than stereoscopic vision, like lidar, will fail in many conditions like snow/rain so you might as well never rely it.
The fact that we can do it with just vision means the computer vision should be able to as well, ideally better.
As a Tesla owner with a model 3 that used to have radar until Tesla disabled it I find this argument funny.
Radar failed when the front got covered in ice, that's about it. With Tesla Vision (camera only), it no longer has any active sensors so it means it doesn't really work and is generally unavailable (as in, not even cruise control is available) or highly unpredictable in conditions such as:
- Rain
- Fog
- Snow
- Low sun
- When using windshield fluid
- Sometimes when driving next to a truck (not sure if tunnels are still an issue as well?)
- Probably more scenarios I forgot about...
Having tried both setup, it's pretty clear to me that the best real-life results are with multiple sensor technologies. I have very little faith Tesla succeeds as long as they are camera-only.
Radar can be supplemental, but it has many false positives. Blind spot monitoring will go off if you pass metal poles a certain way.
You could invalidate that false positive with another sensor, probably computer vision, so you should have just used only CV in the first place.
You have to make sure you only use cameras because you need to drive in all scenarios. If you can't visibly see, the other humans can't either, and it's acceptable to pull over and wait.
Relying on additional sensors just makes your driving less reliable when you don't have it available. Best to only use the common denominator. Less noise, more predictable.
That is Tesla's thinking. If someone wants to prove them wrong, they are free to make an attempt!
When I cross the road I mainly use my eyes, so I should ignore my ears if they hear a car coming because that signal should not be needed.
I normally spot fire with my eyes by seeing the smoke, so if I'm lying in bed at night and I smell burning, I should simply ignore it because that signal should not be needed.
I think the problem with full self driving is that its written from an egotistical viewpoint.
What i mean by that is that when i am driving, i have to take into account the people around me, their vehicle capabilities and more.
For instance you are not going to overtake the car to your right, if you see in your side mirror that there is a junior in a blackened out beamer pulling up in high speed. You will prepare. The AI just sees another car on the right and expects it to behave "normally"
Its either that, or they all communicate intent over wireless tech to each other.
Given that a 16 or 90 year old behind the wheel is statistically more likely to have an accident than a 40 year old...Why wouldn't any sufficiently sophisticated self-driving algorithm be able to make the same statistical data-based conclusion/reaction?
Not going to happen tomorrow but in 20 years? 30 years?
Could see a by then geriatric Musk offering it as an extra add-on "bad driver stereotype" package.
Yes that's true but i mean it more in a sense that you need a complexer model for the things in your surrounding.
basically multiple models for different types of real world objects.
Like when a car has "jackass mode" the other one needs to recognize that car as such. Kind of how we stereotype tainted windows, chrome wheels and AMG lables on the back of a car.
In a "perfect ai data world" the car could query the license plate, compare its actions to previous actions and devise a score just for that car. But it can't do that and hopefully never will, so what remains is a quick computation of the objects in your sourrounding, henceforth a - lot - more computational power
What you are saying is that driving is more than just a technical problem. Driving is a social endeavor. It's a lot about anticipating what other social beings are going to do.
If we have a mix of human drivers and AI assisted drivers on the road, nothing short of AGI is going to be sufficient. Aspects of driving involve theory of mind - anticipating what another driver might do, sometime eye contact, etc.
An alternative solution would be to allow only autonomous vehicles and have these vehicles in a mesh network so can coordinate to drive efficiently and avoid mishaps. This would work reasonably on highways, but on local roads with pedestrians, bicyclists, and unexpected obstructions, theory of mind and ability to handle outliers will be important.
My conclusion is that something close to AGI would be required to implement succesfully.
> An alternative solution would be to allow only autonomous vehicles and have these vehicles in a mesh network so can coordinate to drive efficiently and avoid mishaps.
Why should more public space be given up for private transit? At that point we would be better off building a wide network of autonomous trams with small capacity and high frequency.
21 comments
[ 2.9 ms ] story [ 57.9 ms ] threadI suspect 2d only scanners is a symptom of a deeper problem. It's essentially (premature) optimision for an unsolved problem. And they're underestimating the problem. How long is it before a car can read the cues of a pedestrian that may or may not be about to cross the road without looking, or a dog or a teenager. Has that car stopped because there's something blocking the way, or have they just pulled over without indicating? Is the person just ahead that is driving unusually about to pull into your lane? The long tail of edge cases is just impossibly long.
The fact that we can do it with just vision means the computer vision should be able to as well, ideally better.
Radar failed when the front got covered in ice, that's about it. With Tesla Vision (camera only), it no longer has any active sensors so it means it doesn't really work and is generally unavailable (as in, not even cruise control is available) or highly unpredictable in conditions such as:
- Rain - Fog - Snow - Low sun - When using windshield fluid - Sometimes when driving next to a truck (not sure if tunnels are still an issue as well?) - Probably more scenarios I forgot about...
Having tried both setup, it's pretty clear to me that the best real-life results are with multiple sensor technologies. I have very little faith Tesla succeeds as long as they are camera-only.
You could invalidate that false positive with another sensor, probably computer vision, so you should have just used only CV in the first place.
You have to make sure you only use cameras because you need to drive in all scenarios. If you can't visibly see, the other humans can't either, and it's acceptable to pull over and wait.
Relying on additional sensors just makes your driving less reliable when you don't have it available. Best to only use the common denominator. Less noise, more predictable.
That is Tesla's thinking. If someone wants to prove them wrong, they are free to make an attempt!
I normally spot fire with my eyes by seeing the smoke, so if I'm lying in bed at night and I smell burning, I should simply ignore it because that signal should not be needed.
unless the car is honking it's pretty hard to hear anything due to insulation, especially on the highway.
but yes, humans are very good at using all the sensors they evolved with, the brain is amazing.
maybe one day someone will be able to balance all of their sensors and beat Tesla, or maybe Tesla will figure it all out.
can't wait to find out!
By the way, it's always astonishing how some people can screw up so bad and still come out ahead w/ money and reputation.
What i mean by that is that when i am driving, i have to take into account the people around me, their vehicle capabilities and more.
For instance you are not going to overtake the car to your right, if you see in your side mirror that there is a junior in a blackened out beamer pulling up in high speed. You will prepare. The AI just sees another car on the right and expects it to behave "normally"
Its either that, or they all communicate intent over wireless tech to each other.
Given that a 16 or 90 year old behind the wheel is statistically more likely to have an accident than a 40 year old...Why wouldn't any sufficiently sophisticated self-driving algorithm be able to make the same statistical data-based conclusion/reaction?
Not going to happen tomorrow but in 20 years? 30 years?
Could see a by then geriatric Musk offering it as an extra add-on "bad driver stereotype" package.
basically multiple models for different types of real world objects.
Like when a car has "jackass mode" the other one needs to recognize that car as such. Kind of how we stereotype tainted windows, chrome wheels and AMG lables on the back of a car.
In a "perfect ai data world" the car could query the license plate, compare its actions to previous actions and devise a score just for that car. But it can't do that and hopefully never will, so what remains is a quick computation of the objects in your sourrounding, henceforth a - lot - more computational power
Just think of the times you gladly give away your right of way
Regardless this is an error margin and reactions speed issue ? It’s not like if the car can’t abort if the error margin keeps shrinking
An alternative solution would be to allow only autonomous vehicles and have these vehicles in a mesh network so can coordinate to drive efficiently and avoid mishaps. This would work reasonably on highways, but on local roads with pedestrians, bicyclists, and unexpected obstructions, theory of mind and ability to handle outliers will be important.
My conclusion is that something close to AGI would be required to implement succesfully.
Why should more public space be given up for private transit? At that point we would be better off building a wide network of autonomous trams with small capacity and high frequency.
Be honest at least and quote the figures per 1000 vehicles or something