Lords of Inner Filth

Sep. 9, 2012 at 3:53am with 32 notes

Reblogged from c1qfxugcgy0

The enduring tragedy of the robotics industry is the fact that humans exist.

The Willow Garage PR2, pictured above, has an arm payload of 1.8Kg. (4lbs) That’s… not great. But we’ll give it a pass on that, nobody’s perfect, it’s a brand new product, you have to make compromises somewhere, blah blah blah.

The total robot payload is 20Kg. That’s it. That’s the “combined weight of all additional (non-original/non-standard) components or parts attached to or carried by the robot”. Stuff it’s carrying on the deck, sensors attached to it, whatever: you only get twenty kilos. The PR2 by itself masses 480Kg. It has a maximum, do-not-exceed speed of 1m/s. (2.23mph)

The PR2 costs four hundred thousand US dollars.

In general, technology works best when you give it a hard quantifiable goal to work towards, and no design constraints other than the Eternal ones. (minimize mass use, minimize energy use) Nuclear submarines swim faster than any fish, a 747 flies faster than any bird; you might note that neither of these machines accomplish these feats by wiggling their tails or flapping their wings.

Nuclear submarines and airliners work to the strengths of the materials they’re constructed from, (Aluminum is best at tolerating a very wide range of temperatures, and very bad at automatically repairing itself, or pooping on statues) and do so by adopting radically different forms. If we insisted that airliners weren’t allowed to use internal combustion engines, then flight would either be impossible or very hard.

Since day one, the point of the robot has been to replace humans. In the only field of robotics to see any real success, industrial automation, this has been done by playing to their strengths. They don’t sleep, they don’t have to breathe, you can bolt them right to the concrete floor, most of the ways you can build a servo makes them very repeatable, etc. A welding robot can authentically be better than a human in every way that matters when you’re welding two things together.

Industrial robots accomplish these feats by, generally, looking nothing at all like humans. Most of them don’t even have eyes. But when you introduce the crippling design constraint of anthropomorphism, then you get problems.

The PR2 has to be mobile, which is bad, because there’s no good way to store electricity. There are many kinds of electrochemical battery, superconductor ring, and fancy hydropower tricks; all of which are knocked into a cocked hat by a stone-age conventional gasoline engine. By multiple powers of magnitude. Not even a contest.

So the PR2 will be slower, less powerful, and mass more than a robot which is directly connected to mains electricity.

It also has to be about as big as a human. Strike two: servomotors have many virtues, but myosin has them beat on a power per mass basis. So the PR2’s heavy but weak, as we’ve seen above.

Third strike: Machine vision is hard. The PR2’s head, as seen above, has two sets of stereo cameras, a structured-light depth camera, most of the time a Kinect sensor, and a laser scanner.

An anthropomorphic robot performing human tasks is the absolute worst case scenario, the perfect storm of things that machines are bad at. On one level, this video of a PR2 fetching a beer from a refrigerator is deeply impressive. It is a literally groundbreaking achievement: a robot performing a task that has never been done by a robot before.

On another level, it’s a roomful of engineers spending $400,000 to slowly perform a task that could be better done with a series of tubes. Even with all those cameras, it has to stare at the interior of the refrigerator for several minutes to figure out what it’s looking at. (Note the subtle cheat: the bottles are mounted in a rack dispenser, which significantly narrows the solution space)

The PR2 is a prototype, and as such is a splendid research tool. But even if machine vision was a solved problem (which would probably require strong AI) then it’s not clear if it would be a viable consumer product. It’s not even clear if it would be a good buy if it was a full thousand times cheaper.

I’m tempted at this point to draw a comparison between the PR2 in 2012 and the PDP-11 in 1970. (a rhetorical device I’ve abused before) But not every technology follows the price curve of computers— computing is the bizarre outlier. If you had predicted in 1970 that computers would be tens of millions of times cheaper per FLOP in the future, you would be a visionary, but if you predicted that jet airliners would cost $10, fit in your pocket, and fly at Mach 100, then you would be an idiot.

(However, if you assume mature machine-phase nanotechnology, and a post-scarcity economy, then certain other assumptions no longer hold true)

Anthropomorphic robots are way the hell to the left on the hype cycle. There is not one single component of them that is anywhere near ready for prime time: not energy storage, muscles, computers, or software. Building a better human is a hard, hard problem and after decades of progress, we’re still many decades away from solving it.

(EDIT: Ha ha, and then I see this post on tumblr radar…)