[ExI] self-driving cars again

[Kelly Anderson]

On Sat, Jul 14, 2012 at 10:23 PM, spike <spike66 at att.net> wrote:
>>... On Behalf Of Kelly Anderson
> Subject: Re: [ExI] self-driving cars again
>
> On Sat, Jul 14, 2012 at 12:07 PM, BillK <pharos at gmail.com> wrote:
>>> Slate reasons this is a show stopper for auto^2s: they only go the
>>> speed limit always, and no one wants to go that slow:
>
>>...I don't see why they should not go faster than people...
>
> Besides the still-unsolved problems of legal liability,

Granted.

> note that the
> difficulty of the control problem posed by auto^2 scales by a function far
> more complicated than any known polynomial relationship to the speed.

Yes, that is somewhat the case... but I don't think it's exponentially
more difficult.

> Some
> form of auto^2 existed back in the mid 80s.  When I was up on the range at
> China Lake setting up for an astronomical observation, there were marines
> out there with an automated jeep.  It was going about walking speed.

In the mid 80s (which was the last time I did any work with software
relating to autonomous vehicles, and damn little then) the primary
problem relating to what works today is that we didn't have distance
maps. In today's autonomous world, you have hardware similar to the
Kinect (in functionality, not implementation) that gives you R,G,B and
distance for every pixel. Back in the mid 80s, we were deriving depth
images from binocular vision using optical flow and other
computationally expensive techniques. We still can't really do
binocular vision (it's not really a completely solvable problem
mathematically for various reasons) but we now have depth maps from
hardware.

> Twenty
> years later, I was at the first DARPA challenge, where about thirty entrants
> attempted to herd their auto^2s over a desert road at an average speed of 20
> mph.  All failed, every one.

And this was primarily a function of software that was solved the very
next year when I believe three teams finished. That was in 2005. Now
the software in 2012 is pretty good, at least Google's version of the
software. Thune and probably a lot of the same algorithms as the
Stanford "Stanley" Darpa challenge teams. (Wasn't he one of the guys
teaching that AI course last fall too?)

> I would argue that it might eventually be possible to go really fast with
> auto^2, but the problem is vastly more difficult than it sounds.  There are
> a large number of effects that can be ignored completely at small speeds,
> but which increase dramatically with speed.

While I don't disagree, flowing with traffic also has it's advantages.
If the speed limit is 55, and everyone else is going 65, you have more
difficulty computing how to drive at 55 than 65 because you can't as
easily follow traffic when it is flowing around you. Locking onto the
car ahead of you is a simple enough trick that it would work most of
the time as a default mechanism. It's what people do for the most
part.

> Thune is offering a free class through Udacity on programming auto^2s.
> Check it out.

I will. Fun! If I do that course, then I'll really be able to know
better how it all works today... in more detail.

>>> ... So when you are stuck in the traffic, the masters of the universe
> will be hurtling past at 200 mph.
>
>>...I agree with Bill here.. I think the auto-autos will go much faster,
> though 200 mph might be a bit of an overstatement. -Kelly
>
> I can easily imagine it taking a full decade of proofing before auto^2s are
> allowed to go faster than the current speed limits, and even then, I can
> imagine they wouldn't do it because of the challenges of interacting with
> carbon based drivers.

Once most of the carbon based drivers are gone... then all bets are
off... though you'll always have to deal with older generations of
autonomous cars too.

> Here's what I expect to see very soon: robot racing, probably on a dirt oval
> flat track.  That would be a really cool control problem to try to solve.

I agree that this could happen really fast.

> In flat track, the cars actually have their wheels turned right while they
> go left: the back end is splayed outboard and the wheels spinning.  I can
> imagine something like half scale sprint cars:
>
> http://www.youtube.com/watch?v=gyWwSG5a5gk
>
> so the cars would be about 100 pound affairs with 500cc single cylinder
> motorcycle engines for instance, and could be built for about a couple
> thousand dollars, not including the cost of the software and sensors.  That
> will be a really fun sport to watch.  I predict some form of robot racing
> will show up in the next five years.

I don't doubt it at all. Racing has always been the place that people
have tried out new technologies. So how long will it be before they
allow a robot driver to win at Indy? I say win because if it's good
enough to be allowed to drive, it's going to be good enough to likely
win.

-Kelly