Our eyes are not LIDAR, though, and we can drive pretty well.
In fact, the reason we have crashes is NOT our eyes’ lack of distance detection through laser return timing — having two eyes is enough for distance appreciation. We have crashes because of attention deficit instead.
At this point, there is no reason to believe that a machine can't achieve and outperform a human on a driving task given the same inputs. Sure, human eyes have 5 million cone cells and 1080p feeds only have 2 million pixels, but 4K has 9 million, and more importantly, that level of precision is unnecessary for regular driving.
And Tesla doesn’t even bet just on the visible spectrum; it also relies on radar.
The trick is in our wetware. What the brain does with visual input is not just trivial object recognition. It relies on a complex internal model of the world to both augment the object recognition and to sometimes discard the visual data as invalid.
So sure, theoretically cameras would be enough. But we're not yet there with software, we can't use the camera input well enough. So if you can side-step the need for not-yet-invented ML methods by simply adding a LIDAR to a sensor suite, then it's an obvious way to go.
Compare with powered flight: we didn't get very far by trying to copy the way birds do it. The trick is in the super-light materials birds are made of, and the energy efficiency of their organisms. We only succeeded at powered flight when we brute-forced it by strapping a gasoline engine onto a bunch of wooden planks.
> It relies on a complex internal model of the world to both augment the object recognition and to sometimes discard the visual data as invalid.
That in particular is what makes the hiring fascinating. This problem is Andrej Karpathy’s expertise[0]. His CNN/RNN designs have reached comfortable results, in particular showcasing the ability to identify elements of a source image, and the relationship between different parts of the image.
The speed at which those techniques improve is also stunning. I didn’t expect CNNs to solve Go and image captioning so fast, but here we are!
I think the principles are already there; a few tweaks and a careful design is all it takes to beat the average driver.
Actually in the dark you often drive as a leap of faith in the state of the road. I.e. with very little visibility on what can come from the side of the road (no light) or after a turn. We shouldn't. But we do.
Yes, but it will cause you to slow down. Or at least, it should. As soon as you have less than your stopping distance of space in front of the car you are going too fast.
There's always the option to enable the headlights in such a case. LEDs switch on way faster than incandescent bulbs. And if they're needed at all - a camera has way more flexibility in brightness input range than a human eye.
A car like a Tesla has also highest-quality maps and GPS sensors - these alone are way better than what you get in your smartphone and are enough to keep the car from going over the cliff.
It's pretty obvious that a self-driving camera-based software would take advantage of headlights on a car, just like a human does? So it never has to drive in full darkness.
Not really. Self driving cars should be able to drive in all of those conditions (and one condition might even quickly turn into another, e.g. sunny into rain).
If a technology only helps with some of the cases (e.g. fair weather) and does not work for the others, then there are two cases:
(a) A single replacement technology will be found that works in 100% of cases.
or:
(b) The technology will only be used on the cases it works well, and the other cases will be handled by some alternative technology equally only suited to them.
In the case of (a), Lidar is indeed useless (or at best, only used as a supplementary technology in favourable conditions).
And I fail to see how (b) can be the case -- that is, how there can be another technology that will solve the rain/snow/night driving problem, but which cannot also outperform/replace Lidar for fair weather driving.
Isn't it interesting that we have five senses, when we could just have one that works in 100% of the cases? A third option is a system based on multi-sensory inputs. Several inputs that are just marginal on their own can provide good performance when combined.
Our eyes are not LIDAR, though, and we can drive pretty well.
In fact, the reason we have crashes is NOT our eyes’ lack of distance detection through laser return timing — having two eyes is enough for distance appreciation. We have crashes because of attention deficit instead.
At this point, there is no reason to believe that a machine can't achieve and outperform a human on a driving task given the same inputs. Sure, human eyes have 5 million cone cells and 1080p feeds only have 2 million pixels, but 4K has 9 million, and more importantly, that level of precision is unnecessary for regular driving.
And Tesla doesn’t even bet just on the visible spectrum; it also relies on radar.
The trick is in our wetware. What the brain does with visual input is not just trivial object recognition. It relies on a complex internal model of the world to both augment the object recognition and to sometimes discard the visual data as invalid.
So sure, theoretically cameras would be enough. But we're not yet there with software, we can't use the camera input well enough. So if you can side-step the need for not-yet-invented ML methods by simply adding a LIDAR to a sensor suite, then it's an obvious way to go.
Compare with powered flight: we didn't get very far by trying to copy the way birds do it. The trick is in the super-light materials birds are made of, and the energy efficiency of their organisms. We only succeeded at powered flight when we brute-forced it by strapping a gasoline engine onto a bunch of wooden planks.
> It relies on a complex internal model of the world to both augment the object recognition and to sometimes discard the visual data as invalid.
That in particular is what makes the hiring fascinating. This problem is Andrej Karpathy’s expertise[0]. His CNN/RNN designs have reached comfortable results, in particular showcasing the ability to identify elements of a source image, and the relationship between different parts of the image.
The speed at which those techniques improve is also stunning. I didn’t expect CNNs to solve Go and image captioning so fast, but here we are!
I think the principles are already there; a few tweaks and a careful design is all it takes to beat the average driver.
[0]: http://cs.stanford.edu/people/karpathy/main.pdf
2 replies →
Correct, eyes in fact are useless if you can't see in the dark.
Now imagine walking on top of a sky scraper in pitch darkness. Yes your eyes work in light, but in this case you will likely fall to death.
You can't really drive in the dark, can you? What if it gets dark for 1 second on a cliffside turn?
Actually in the dark you often drive as a leap of faith in the state of the road. I.e. with very little visibility on what can come from the side of the road (no light) or after a turn. We shouldn't. But we do.
Yes, but it will cause you to slow down. Or at least, it should. As soon as you have less than your stopping distance of space in front of the car you are going too fast.
1 reply →
There's always the option to enable the headlights in such a case. LEDs switch on way faster than incandescent bulbs. And if they're needed at all - a camera has way more flexibility in brightness input range than a human eye.
A car like a Tesla has also highest-quality maps and GPS sensors - these alone are way better than what you get in your smartphone and are enough to keep the car from going over the cliff.
It's pretty obvious that a self-driving camera-based software would take advantage of headlights on a car, just like a human does? So it never has to drive in full darkness.
You missed the analogy:
Darkness is to eyesight as inclement weather/obstacles is to LIDAR
3 replies →
Actually you will need more than a camera input, may be even a sound input.
In darkness it helps if you can hear the vehicles coming close.
2 replies →
Not really. Self driving cars should be able to drive in all of those conditions (and one condition might even quickly turn into another, e.g. sunny into rain).
If a technology only helps with some of the cases (e.g. fair weather) and does not work for the others, then there are two cases:
(a) A single replacement technology will be found that works in 100% of cases.
or:
(b) The technology will only be used on the cases it works well, and the other cases will be handled by some alternative technology equally only suited to them.
In the case of (a), Lidar is indeed useless (or at best, only used as a supplementary technology in favourable conditions).
And I fail to see how (b) can be the case -- that is, how there can be another technology that will solve the rain/snow/night driving problem, but which cannot also outperform/replace Lidar for fair weather driving.
> then there are two cases:
Isn't it interesting that we have five senses, when we could just have one that works in 100% of the cases? A third option is a system based on multi-sensory inputs. Several inputs that are just marginal on their own can provide good performance when combined.