But today, I've come to teach a thing or two.

To the robots, that is, not the people.

In a bid to help them behave as humanly as possible, especially with the tasks that we wouldn't give a second thought to.

This reinforcement learning where the machines mimic humans is how the AI is trained.

It aims to make the bots less, well, robotic and more multi-skilled.

This motion is actually independent of where we place the glass and the picture, so we can have a much more adaptive system that does not need to have everything specified precisely.

In fact, this is part of the magic of machine learning.

With machine learning, we can turn pre-programmed robots into intelligent machines.

That means we can use data, we can use text, we can use images.

We can even use muscle activity, and we can learn from this data how to do tasks in a much more seamless way.

Baxter's also been trained on doing the dishes and chopping veg.

Wouldn't that be nice?

Meanwhile, robotic hand, anyone?

So, Laura, this is another robot we've been working on, and what's special about the hand is that it's got a very compliant body, so it's got skin, in other words, soft material.

Can I touch it?

You can touch it.

It feels like silicon, but I can also feel what is like the bones inside.

So the reason we put bones inside is because we want to have a hand that is compliant, and that's what we get from the silicon.

On with the glove, and time for a lighter task to start.

My hand's starting off in the same position as the bot, and the idea is that I'm going to train it to pick up that dollar bill.

It's really sensitive to movement.

As you can see here, just the tiniest turns, and you can really see them.

Also, those movements are quite human.

If I do a kind of typing action, look at that, look at the fingers in real time.

Right, so if I now move my hand down as though I'm going to pick up the dollar bill...

OK.

You're off, you're off, you have to shift.

I tried and failed repeatedly, so I'm not sure how much use this training was.

Oh, so close, it's so frustrating.

Next up, a look at how robots can work as a team.

Think swarm of ants.

What sort of real-world uses would robotics created on this basis have?

You could take a swarm of robots like these ones and get them to go up a bridge to ensure the integrity of the bridge.

We have to consider what should the body do, what should the brain do?

And depending on the task of the robot, how should we design the body so that the body is capable of the task?

But that's not enough.

We then need the brain to get the body to do what it's meant to do.

And I suppose the interaction between devices is really important for real-world purposes like self-driving cars, that they be able to see each other and understand how to interact with what's around them.

Exactly, so the coordination and group behaviour we get from local interactions to global behaviour are super important.

That is part of physical intelligence.

That is part of getting robots to perceive the world around them and reason about it.

And now it's turtle time.

Krush is autonomous, having learnt to swim from data collected on real sea turtles.

This actually gives us a very naturally moving robot that goes in the sea and does not disturb the sea life and that allows us to study the sea life better.

This robot is able to move whatever way we want in the sea, all of this enabled by its algorithms.

I can actually see how natural the movement seems.

The fact that it's not that even.

It's smooth, but it doesn't seem that sort of organised in its movement, which is, of course, a very real way that a creature would behave.

Exactly.

From the seas to the skies, devices are not only learning from the physical world, but also from language and AI's ability to reason.

Move towards the bag.

This drone is being trained to understand instructions in any language.

It's never seen a red bag or a traffic cone in its data set, but combining its understanding of language and vision, it is able to reason, so should be able to understand and carry out this command and so many more.

And there it is.

Just by me instructing it in my own words, it ended up in the right spot.

Before I go, though, I have been promised some lunch.

Someone's making me a sandwich.

No pressure, but I am quite hungry.

Not sure how appealing that lettuce is looking after having that done to it.

It's not manhandled, it's robot handled.

It's getting there.

Turns out that for robots, some of the jobs that we find the easiest are going to take some time to learn.

It's like watching someone with a claw machine, just a really advanced version.

And there we have it.

Thank you.