Original subtitles

Speaking of robotics, let's talk about what we can do in robotics. We cannot yet

send them through time.

But we can maybe do more than you may think. So what is robotics? Well to

actually build robotics there's a lot of say mechanical engineering,

the mechatronics, building the gears and and sorting out the forces and soldering things.

That's all very important,

but it's not really what we do in this class. In this class, we look at

the part of robotics which is AI.

And that's the control of the robots, figuring out,

once you've got this robotic platform,

what control signals do you send down the wire to do whatever you're trying to do--

pick up the cup, or sweep the floor, or whatever it is.

It turns out that simulating robots is much easier than actually

deploying robots. We alluded to this earlier were once you're in the real world,

suddenly it's not just planning for how you move that arm, but it's acknowledging that

maybe your sensors are a little bit wrong or maybe the arm will slip during

motion, and everything is much, much harder in reality than simulation.

Also when you mess up it's more expensive in reality than in simulations. These robots

do not come cheap.

What kind of things we build? We can build autonomous vehicles. We'll talk about a couple kinds in this course.

We can build rescue technologies, we can build soccer-playing robots which is

of course a critical goal of AI, and we can automate a bunch of stuff. I mean in some sense there's

robots all over. Your dishwasher is a robot for cleaning dishes. It's just,

to a first approximation you blast water at them and they're clean, so we

often don't think of that as much of an AI task,

though there are control problems even in a dishwasher.

In this class we're going to ignore the mechanical and mechatronic aspect of robotics

and we're only gonna think about the planning and control aspects.

Let me show you for today, a kind of sense of what kinds of things you can do.

I love this. Does anyone know what this is?

This is RoboCup, this is the Aibo league. They're so cute.

I'm going to play a little bit of an Aibo soccer match.

The first thing to notice is that

they're in formation,

they're going to play soccer. It's a team thing; they have to coordinate.

It's kind of like watching five-year olds play soccer right.

Like the white goalie, what are you doing,

what are you doing.

Fortunately you don't actually need a goalie when you're playing against these guys.

Goal!

So, what to learn from here. First of all multiple agents are harder to control than one--

it's actually hard to do much with an Aibo--but, multiple agents are harder to control.

You've got to coordinate. This this level of multi-agent control.

You also have to get the individual robot to simply walk along, and like find where the ball is,

do the vision and the calibration.

All of that stuff is hard.

Even when you're not thinking about soccer strategy.

And so this is an example of that.

Something else I'd like to use these

Aibos to show is that remember we talked about the difference between

an AI acting in an optimal way

vervus like a biomimetic kind of human-like--or in this case I guess dog-like--way?

It turns out, when you're playing soccer, you know, a dog might move a ball a

certain way. But if you've got an Aibo,

you don't want to move the ball like a dog.

You want to move the ball in the best way possible. And you'll never guess what it is.

And then you've got to turn around and see what the heck happened. So they're all going to

go over to the ball now.

So, again, the point here is,

optimal,

sometimes you kick it the other way.

Not biomimetic.

We talked about autonomous cars. So here's some footage from a Google autonomous car,

and you'll notice,

it may not be Telegraph Avenue, but

these things are kind of around you driving and the cool thing is

you probably don't even know it. Sometimes you

look over into the next car and the robot waves back,

but you know barring that you may just not even know.

Fully autonomous driving.

This is showing some of the sensors that are available.

I don't know, like my palms sweat while watching this driving even

with a driver there right.

So, autonomous driving,

better than you might think.

We talked about towel folding, so

let me show you

a bit of towel folding from Pieter Abbeel's lab.

This is the PR2,

folding towels.

it turns out when you're a robot,

the hard part about folding is not the folding part, but in the figuring out

where the corners are in the first place. So you'll see the robot kinda like

twirls the towel until it finds a corner, and then once it finds a corner, boom boom boom.

well, you know, it's sped up 200x so it's not boom boom boom, but

booom, booom, booom.

So it folds them, stacks them,

beautiful.

It is currently not the most cost-efficient way to fold towels.

Okay now let me show you the bad boy here. So this is a

Boston Dynamics robot.

Basically if Terminators had really really tiny heads,

they'd look something like this.

So let's see what this thing can do.

It can pose.

This robot, it has kind of a big range of motion.

It can be stable doing these rotations which is amazingly hard--people do it.

It can walk.

If I saw that walking towards me I might be a little worried.

it can strike its pose and kneel down and do things. It does push ups in its spare time!

Okay you know when your robot does pushups for fun,

you've got a certain kind of robot.

So, really amazing stuff. It may not be obvious from watching this but it is amazingly

hard just to get a robot to walk. That why many robots have wheels,

an easier platform of some kind. To be able to walk, like this, this is amazing stuff they have.