Tesla has a level 4 autonomy demonstration video on their site. Maybe computer vision is superior to LIDAR; not a lot of human drivers with laser arrays on them driving around.
Yes, the software is the hard part. That's why it makes sense to use hardware that makes the job of the software easier. LIDAR tells you how far away objects are; if you're using cameras you have to do that with software instead. AFAIK no one yet knows how to do computer vision with cameras in the sorts of situations a self-driving car needs (though I'm not a vision researcher). That means that Tesla is betting on solving a research problem to get the cars they've already sold to start driving themselves. I always thought that was an insanely audacious business plan.
I don't think this term even means anything. The problem can be solved by various mixtures of hardware plus software to varying degrees of adequacy.
On the one hand, we know that it's possible to drive a vehicle with high reliability in a wide range of circumstances with just two relatively low resolution cameras... if those cameras are feeding into a trained human-level intellect. What combination of hardware and software is required to generate that intellect is still unknown.
On the other hand, the problem seems relatively well solved using 3D laser scanners and relatively straightforward software. It works pretty well and just needs to be shaken down in all of the possible circumstances it might encounter. It's also unreasonably expensive.
We seem to be converging on an "easiest" solution which uses a mixture of cameras, radars, ultrasound and maybe LIDAR, using deep learning for image processing. That's just one solution, though.
We probably can, but most of us drive with two cameras with approximately 120 megapixels of feature resolution with highly optimized non uniform pixel density and incredible dynamic range. Micromovements allow the software to boost the effective resolution by about a factor of four, to over 400 megapixels. Image stabilization works in multiple layers and executes stabilization movements within 10ms in the fastest layer.
The human eyes and vision system are certainly one of our more impressive parts, and we have a really hard time replicating their performance.
Lots of LIDARs, lots of cameras.
"Additional guidance data was provided by a video camera used to observe driving conditions out to eighty meters (beyond the range of the LIDAR) and to ensure room enough for acceleration."
"Stanley was characterized by a machine learning based approach to obstacle detection. Data from the LIDARs was fused with images from the vision system to perform more distant look-ahead. If a path of drivable terrain could not be detected for at least 40 meters in front of the vehicle, speed was decreased and the LIDARs were used to locate a safe passage."
This computer vision processing was performed with 1.6ghz Pentium M processors 11 years ago. Is it so far fetched Nvidia parallel processing hardware can't do better with CCDs than LIDAR?
2) In theory almost everything is "just" a compute problem (just enumerate or search over all possibilities - of course this may not finish for millenia). But if Moore's law says the compute is 20 years away, I think it's reasonable to call it a research problem.
The 100k car/driver thing is a red herring. Steering angle / brake / acceleration info is not the kind of data you need for this task. Big useless data is still useless data, even if it's big.
The right kind of data, if this is to be treated as an ML problem, is a 3D model of the environment (this would be the label), along with camera data (this would be the data). Ironically you need LIDAR to get those labels. It bears noting that the state of the art in visual SLAM is not ML based (it's really difficult to get the "ML" problem I just described to actually work - and it's not for lack of data, Google probably has petabytes of this data from their cars).
Those are labels for a different problem.
Not to say LIDAR isn't essential. I think LIDAR will get cheap enough that it'll just be silly not to have it.
IMO the most successful implementations are ones that can predict accidents that humans couldn't even see taking place (obscured by other cars, etc).
Such as Tesla'a forward facing radar, highlighted the other day on HN preventing an additional accident participant by detecting a vehicle slowing down ahead that the driver could not visually detect?