I. The rise of robots

Ever since I was a kid, I’ve wanted robots. I'm not alone in wanting this either. Robots are the basis for any utopian vision of the future. And as we're starting to see this future coming together, with AI able to write emails and code and make art, this is the part that's still missing.

Yes it's hard, and yes we've tried in bits and spurts with limited success, but many many things are hard and we do them. So why not this? All we're left with is a lonely roomba getting stuck on the odd carpet or litter box.

When I tried to ask why this isn’t here yet, the major objections were:

The cost is too high

Real world is too hard to navigate

Hardware isn’t ready eg tactile sensors

Battery density is too low

This devolves to a function of two problems - they don’t work very well, and they are not affordable.

Let’s have a look.

They have to work

One of my favourite things I did was to teach a robot how to walk, back in 2005. At that point, since working AI was a pipe dream, you had to think about the mechanics of the joints, angles of bending, weight distribution and centers of mass, inclines of the path ahead, and a whole bunch of other things to be able to create some boundaries within which the robot learns, to train its neural net.

(Parenthetically one part I remember is naming a bunch of variables after characters in the Prince of Persia videogame, just because we were incredibly cool in college.)

And turns out that this is incredibly hard. There is no way you can handle all edge cases when the training is pretty specific, and rooted in biomechanics. To be successful, they have to work relatively autonomously, navigate its surroundings, make autonomous decisions, and be able to actually handle things like a baby blanket or a hot cup of coffee.

To actually do them in real life requires some form of reinforcement learning (I wager) to get the inner equilibrium etc right. I assume this isn't trivial, though I've only done it the hard way. But OpenAI did set a benchmark for robot dexterity, and later abandoned its research because there was no large enough dataset to help power its reinforcement learning.

And the robots of today seem pretty similar. Boston Dynamics have some of the best videos of robots doing cool stuff I’ve seen, but their training method seems awfully similar to ...

Posts

Wiki Contributions

Comments