Autonomous Systems

Well, it's not required persay. If it was we feeble humans and our addiction to the visible light spectrum wouldn't be able to drive.

It does, however, make things a lot easier.


The granddaddy vehicle lidar, the Velodyne HDL64-E produces a scan with 64 vertical layers. I believe the below is from a single scan:

Here's a google car visualization (somewhat old)


There's two big reasons why lidar is easier than monocam Computer Vision (CV):
1. Going from processing mono-cam images to 3d using structure from motion, edge-searching, object recognition and a plethora of other techniques to lidar like the above is like cheating. You know how far a given thing is from the lidar because you've got a discrete and independent distance measurement.

You aren't susceptible to a bunch of corner cases like your shitty assumptions that worked fine until some jackass in a semi pulled in front of your Tesla and your CV classified it as a harmless billboard overhead (That's what happened in that fatality).

2. Cameras are highly dependent on lighting conditions. Lidar fails in certain conditions too. Ex: direct sunlight in the sensor, isn't a big deal really. Heavy dust obscurants (really big fucking deal that nobody is admitting yet. OPAL LiDAR is trying to solve that via a class 3 laser). Snowfall and ground covering is a huge problem that Ford is trying to solve. But over all lidar shows up with its own energy source and works in a variety of lighting conditions including dark, heavy sunlight. If you ever see a CV example done in an overcast sky, they're probably hiding that it fails in bright sunlight. So why can I drive at night? Well, I use a ton of inference and assumptions. Everytime I turn left around a car with its lights on I'm betting that there's nothing there, but it's dangerous and if there was something there it'd be dead. It'd be trivial to set up a test case and people would run over your black hoodie-wearing mannequin everytime. Hands up don't manslaughter. When I drive into the sun and can't see shit, I'm just hoping that nobody gets in my way. If that light I can't see just turned red, buckle up everyone.

A third less tangible reason is that lidar tech is absolutely exploding right now. Once/If we get Quarnergy style solid-state lidar, it's going to be fucking awesome. I'm hoping for a similar increase in mon-camera tech that more closely matches the human eye in terms of robustness in lighting conditions, but we're just not there yet and that problem is really hard.

ZyyzYzzy said:

Maybe I'm naive but I think your first point is bullshit. How is it cheating? That is just an asinine statement. The visual systems that has resulted over 3 billions years of evolution operates by making inferences like you stated, it is probably more than sufficient for almost every application. Very few animals use anything but in a terrestrial environment, and to my knowledge outside of light void environments none use another sense exclusively. Processing information in such a manner is just as good, as the system is nit overburdened with "useless" information. Animals with really specialized visual senses do pretty amazing things in regards to tracking in the environment (see predatory birds being able to account for refraction of prety in water).

Click to expand...

I don't even know what point you're trying to make here, but when one spends a tremendous amount of effort trying to get structure from motion of objects in a mono-cam, it feels like cheating when one switches to a lidar like the velodyne above that just tells you what everything looks like in 360 degrees with a huge vertical field of view.

I'm saying that it's much easier to perform that kind of processing from a point cloud produced by a lidar than a sequence of mono-cam images. Cheating is good in this case.

If you're trying to argue that because I've got 3 billion years of evolution that have provided me the ability to see well enough with one eye that vehicular autonomy is possible with a mono-cam, I agree.

It's just a lot harder and the human eye has tremendous capability that today's cams don't.

ZyyzYzzy said:

Arguing your initial point that not using LIDAR is near impossible/retarded for anyone to pursue?

Edit - you last point...what? Cameras have far superior abilities to any fucking eye, the difference is the ability in processing the image, which isn't done by the eye. Again if a fryit fly with a few hundred/thousand neurons can track in a complex environment, then developing the ability to do that is not an impossible feat.

Click to expand...

So when you read:

Tuco said:

So, what's great and crazy about Tesla is they're trying to accomplish full scale autonomy without lidar, which is the foundational sensor for everyone else's perception. It's kind of like trying to be a raiding guild by using bards and rangers for tanking. Maybe it'll be as forward looking as having your rockets accomplish landings, but either way it'll be tough to get here.

One important thing about the video is that it was taken in excellent lighting conditions. It was a little rainy, sure, but for mono-cam work an overcast sky with heavily diffused lighting is ideal. Darkness sucks, of course, and bright sunlight can eviscerate your CV algorithms because it'll blind your camera, create dramatic shadows on everything you're trying to detect etc.

The guy running Tesla's autonomous division knows what a lidar can do and I'm pretty pumped to see if they're able to make it work with radar + mono-cams. My expectation is they'll lead the pack on vehicular CV, but once we get solid state lidar they'll adopt it and be in good shape.

Click to expand...

you saw me say "This is great and crazy, and I'm pretty pumped to see how far they get." you interpreted it as, "This is near impossible and they're retarded."?

Go fuck your strawman on your own then.

Tesla making world-class vehicular CV and then switching to lidar once the lidar tech reaches new pinnacles is a totally worth endeavor and might even be a better long-term plan. Vehicular autonomy is always going to need amazing CV, so them pushing it while literally everyone else is pushing lidar might make a lot of sense.

Motorcycles are a significant case that autonomy has to solve, but I expect that it's one of the many cases where autonomy will be able to track motorcycles with much better accuracy than human drivers will, even if it's a problematic case today for Tesla or whomever else.

Motorcycles are tough for human drivers because they are small and can hide in blind spots easily (among other reasons). Autonomous cars have much better 360 coverage. Bikers will always be the #1 person taking care of themselves through very defensive driving, but as vehicular autonomy gets better they should breathe a little easier being near an autonomous vehicle.

Yep, and that's just one of the many behavioral challenges that autonomous vehicles have to surmount to be production ready. And it's easy-mode in the US compared to other countries. In places like India motorcyclists can just be modeled via fluid dynamics given how little they observe any laws.

Vehicle to vehicle communications are one of the big parts of vehicular modernization and will be a big part of autonomy. However you'll probably never be able to rely on it (especially with motorcycles for a few reasons), and even if you could it's unlikely that a vehicle will ever know where its absolute position with more accuracy than another vehicle will know its position relative to it. In other words, a bike may know where it is globally within 2 meters, but an autonomous car will be splashing it with lidar and know where it is (relative to the vehicle) within 2centimeters.

So what about the hacking angle?

Or someone systematically trying to throw wrenches into the digital system?

What about power failures? Ho will those be negated? The more complex a system gets, the more problems it encounters.

Whose to blame when my antonymous vehicle kills someone due to one of the above?

mkopec said:

So what about the hacking angle?

Click to expand...

There's a tremendous amount of focus on security and it's a huge issue but I think it'll end up being easier to hack vehicles by dropping a boulder off an overpass onto them.

Somebody actually did that to my aunt and it was inches away from killing my cousin when he was a baby.