- Robots have been around for a long time.

Lately, they can mow your lawn, drive you to work or school, even clean your house, all without sleep.

But there's a new question now.

Not how long a robot can work, but how far it can think.

What would happen if you could just give a robot a goal and it would figure out how to do it on its own?

- What skills do I need to learn to create this world?

What if we could use code, do things in the physical world, move objects around and assist in tasks that humans don't want to do?

- Where do you wish the shielding stacked, sir?

- Angelous Angelopoulos is a member of a new generation of computer scientists who are helping robots map spaces, recognize objects, and adapt to a world that keeps changing.

Meet Fetch.

So I'm in its way.

So then it comes close to me, it's going to stop.

Because it's like, hmm, there's an obstacle.

- It can do all this in a lab by itself?

- Yes.

- It can find out where it's going.

- Correct.

- It knows what each item in the lab looks like.

- Yep.

- And basically almost be a lab colleague, right?

- Yeah.

And it's not going to hit anything.

- That's just amazing.

There are colleagues, but then there are just robots like these in an auto assembly line, for example, which are built to do one thing every minute of every hour of every day.

- Fetch is different.

It was built for more small environments, indoor environments, and it was built to handle uncertainty more.

- Fetch is a research robot built to move, see, and think, rolling through real spaces, lifting real objects, and learning to work in the same unpredictable world we do.

Sixteen years ago, Ron Alterovitz left Caltech for UNC Chapel Hill.

Now he leads the university's robotics lab, pushing machines beyond simple instructions towards a deeper understanding of the world around them.

- So the lab automation robot, when we bring it into a new space, the first thing it can do is map out its environment.

It can move through it and then create a map of the room, including where it can navigate, as well as all the different instruments that are in the lab.

- Recently, Fetch went to the chemistry department to do some work and demonstrated how precise its work can be and how it might actually work alongside its human colleagues.

- We've also developed algorithms that allow the robot to do specific tasks.

So one example task that we've implemented for the robot is that it can use a needle to pick up a sample from one particular instrument and then transport it to another part of the lab and then inject that sample into another instrument.

And these types of tasks are challenging because the robot has to both move through the environment and then perform a highly precise manipulation.

- Its seven-motor arm is programmed to reach, lift, and place objects where they need to be, with the help of a group of four cameras.

It also uses LiDAR technology at the bottom, which shoots lasers out to determine distance.

And taken all together, it can sense depth in a 3D space.

- This is the primary component.

Here is also a color camera.

But this time it's a much more wide-lens camera, so it can capture a much larger field of view compared to the default one.

And it also has 4K resolution.

So this is very useful for perceiving small objects with high fidelity.

- When we interact with a computer, we type something in, we see something back on the screen.

With the robot, the robot can actually change the state of the physical world.

- So looking at this, it looks a bit almost like Fortnite, like a terrain.

Can you talk about this?

I mean, this is what the Fetch is seeing here?

- Correct.

Yeah.

So this is the map of the room that you see here based on the laser sensors.

These black pixels define where the walls in the map are.

But then you can also see these little colored dots here.

And these are essentially where those laser beams from the base hit.

And you can see they're aligned with the walls.

So these are live data, and they match our map, which is a good thing.

So these also help the Fetch identify things like humans in the space.

Like if we look at other little dots that you see here, this is some person's legs as they move around.

- Is that Ron?

Are those Ron's feet?

- Yes, correct.

And you can see the robot can identify where they are and kind of create these gray blobs around them to mark them as obstacles that it should avoid.

So the basic thing is the robot, in order to navigate around any space, it has to know what that space looks like.

And it has to know where its destination is.

The way this works is when we take the Fetch to a new location, we essentially have it move around and create a map of where all the obstacles are and where is the free space that it can occupy and move through, like the hallways of the lab.

So essentially, if we need to go to a specific instrument, we say, "Go to this location and then look for essentially what looks like a QR code."

And that QR code essentially allows it to locate exactly what the instrument is.

- These QR codes aren't like what you would see in a restaurant that point you to a menu.

These tell the Fetch exactly what the object is.

In this case, it's a gas chromatograph of the UNC Chemistry Lab, which is a machine that analyzes chemical compounds.

- From here, I can essentially have some offset of kind of where this target is, to move the arm kind of in this location.

This camera would see where that needle is, detect it, and then tell the robot, "You're off by this amount.

Move that way."

So this and this are kind of used together to get that certain needle inside this portal.

These are also used for space applications.

So let's say we want to send a robot to the moon and it has to assemble some shelter for a person that will come after the robot.

Well, you can imagine you can put a lot of these on the objects to kind of identify where those holes for the screws should go, so the robot can locate them.

- So even the non-human construction of a space station brings that.

- Correct.

- For now, these machines can only do what we tell them.

The responsibility stays with us, guiding what they learn, laying down guardrails, and making sure they serve our best interests, not the other way around.

- Ultimately, the robots of today are still running programs that were written by humans, and so we have some level of influence on what our robots can do.

- So it really changes the game from capturing really high-quality datasets that have the behaviors that we want the robots to learn, and ensuring that as we scale up those models and make them more intelligent, that they remain aligned with what we train them to do.

But that requires effort.

[laughter] - Great.

♪