Hacker News Reader

Comment by serioussecurity | Hacker News Reader

I keep seeing this meme about Tesla and it keeps making me grumpy. Basic lane keeping is simple, and Tesla is only doing that. And they're doing it badly.

Tesla's autopilot is a PR move to try to raise more capital. They are not and will not be a serious player in driverless cars.

What they've got is a system doing object detection. Object detection is actually one of the easiest parts of driverless cars. Planning and interaction end up being much harder and requiring a lot of data, none of which Tesla has collected. (Also, in practice you need lidar and a fully mapped road, neither of which Tesla has; we're at least a decade from the top players who are actually collecting large scale urban / suburban data from dropping the map requirement, and there's really no reason not to use lidar as the price is dropping so quickly. Both reasons Tesla's "we've got lots of data" claim are BS...)

View on HN

> you need lidar and a fully mapped road

To play devil's advocate - somehow I drove myself to work today, and the hardware I'm running is just two moderate resolution limited field of view cameras. Not an expert, but from first principles it should be possible to pilot a self-driving car with cameras only, given enough processing power and a smart enough agent. Maybe those last two aren't there in 2018 though.

The problem is your "smart enough agent" and how to obtain one.

There is something in the middle between your pairs of eyes and your decision making that makes you understand what is actually going on around you: perception/cognition.

If this was so simple as you put it (and in particular just by throwing algorithms and processing power at it), several problems would be trivial by now and we all already would have our personal digital assistants... I mean the real deal.

Well knowing I sound like "dismissive grampa" right now, and yes, we have come far and it is impressive, but I sometimes feel like us nerds/hackers/software guys tend to considerably underestimate most problems, and as consequence the intelligence and efforts of those that came before us.

> To play devil's advocate - somehow I drove myself to work today, and the hardware I'm running is just two moderate resolution limited field of view cameras.

This sounds like a fallacy (not sure which one). Just because you can do it does not mean a machine can. There are things that babies can do that machines can't (in 2018).

That misses the point. The idea is that a human being is machine running on couple hundred Watts, twenty of which are spent on compute and sensing. There's no reason why a man-made device couldn't replicate the feat.

There is, in fact a reason we can't replicate that today, and it's not for lack of trying. We don't have neither the full understanding of the human machine nor the technology to replicate it. As an example, muscles are well-understood, but we haven't been able to make artificial muscles with energy efficiency in the same ballpark. The human brain is much less understood.

I'm not saying it won't be possible at some future date after some hypothetical breakthroughs, but we are far from it presently.

That is exactly what serious security said - "in practice you need lidar and a fully mapped road". The word "presently", "today" or "with current computational limitations" are notably missing from that statement. It would at least add some ambiguity to a statement that is provably wrong.

This is pedantry. It's the same as them saying "You need a rocket to send a payload into space" and someone retorts "No: you can also use a space elevator". The presently is implicit, and space elevators don't (yet) exist.

The initial statement is saying that there is no physical way to achieve this feat today without lidar. While millions of agents (humans) are currently doing this today without lidar, or even eyes in the back of their head. I don't see anybody/anything with the blasé regularity of human drivers without lidar (machine or not) moving things into space except via space craft. So to boldy claim that lidar is NEEDED is absurd. Speaking of getting things to space, I remember people telling the same guy, Elon, that you NEEDED a new rocket to launch your payload into space every time. It is funny, you needed that until you didn't. All the while, humans we, with the blasé regularity of human drivers without lidar, not throwing away their cargo aircraft every time freighted between airport.

Keep the "pedantry" name calling out of this.

Keep in mind that a self-driving car will necessarily need to have a much safer track-record than a human driver. That's what drives the need for more than just two low resolution cameras looking out the front. Plus you need the cameras to be able to look behind and to the sides of the vehicle which is easier to do with more than two front-facing cameras.

Your eyes have significantly more dynamic range than even the best generally available cameras. For example, your eyes can see enough details to walk in a dark room at night while cameras would need additional illumination (flash or IR).

Smart enough is also pretty hard, especially for edge cases. For example, imagine driving on a highway and there's a discarded grocery bag flying around. Based on the flying pattern, it's pretty easy for you to identify such an object if you've seen one before. For a deep neural network, if it hasn't seen enough examples of such objects, it'll fail to classify it properly. What will a self driving car do if it sees an unknown object in front of it while driving at highway speeds?

So there's a couple things at play.

Lidar range is more than double visual range, in practice.

When figuring out collision avoidance paths for an object you actually end up approximating some np hard problems to find a path that won't have collisions and won't be too "careful".

This ends up being fairly computationally intensive, and adding the extra time significantly improves your planning. Doubling compute time tends to beat doubling your training dataset in terms of system quality, at this scale.

Extra time also turns a number of situations from "guaranteed kill" to "we can avoid the accident", because the car is traveling really fast and those extra seconds can be used to brake, find a new path, etc.

In visibility impaired situations, lidar and vision have different constraints and ways they fail, and the intersection of the two can significantly improve scene understanding ( see waymo's snow demo ).

In a lot of cases, path planning can be dramatically improved by having maps. If you're going into a curve and know the shape of the road, you can preload that and spend your time on more important tasks like object detection and path planning.

Etc etc etc.

This is absolutely not a domain for intuition and thought experiments. The pragmatics of the industry are highly intricate and responsive to constraints that are only visible if you've worked on this stuff.

That's not the only hardware you're running though: those two cameras are connected to a sophisticated object-classification system with around 540 million years [0] of R&D behind it.

[0] https://en.wikipedia.org/wiki/Evolution_of_the_eye

What happens if two cars with LIDAR drive toward each other? Isn't this one of the problems google self-driving car had, where one car interfered with others?

Reply on news.ycombinator.com