[Host] This is the result of your

online shopping addiction.

It's a picker robot.

Its job is to sort through goods

in an order fulfillment center.

A task that's traditionally been difficult

for machines.

But that's changing.

Thanks in large part to a mix of AI

and some good old fashioned human hand-holding.

I guarantee you take for granted

how many things you can pick up.

What comes so naturally to you is

difficult for a lot of robots.

But not these picker robots

at a startup called Kindred.

They're able to grab a variety of items

and hold them for a barcode scanner

and file them in the right cubby hole.

The big opportunity in e-commerce

is that there are millions and millions of

different types of objects.

Packaging keeps changing,

some are soft and squishy,

some are hard, some are heavy, some are soft.

And there's no way you can program that.

The challenge for us is

constantly learning all these new objects

fast enough to respond to our customers.

[Host] But fear not.

The robots still need our help

learning how to tackle these new objects.

The team at Kindred lets the robot

try and grasp objects on its own.

This is known as reinforcement learning.

Meaning, it gets a digital thumbs-up

whenever it does something right.

And adjusts its behavior accordingly.

Humans also refine the robot's skills by

remotely piloting the machine.

[Man] The human controller guides the arm and the gripper

to pick up the objects,

and we use all that data from the gripper

from the servers from the arm

to feed our algorithms.

So the next time they see the same shape,

we know how to pick it up.

[Host] This approach provides a flexibility

that's essential in a job like order fulfillment.

Not only does the robot need to know how to

manipulate a galaxy of different objects,

it has to adapt on the fly to

novel products.

So let's say that coats come into season

for a retailer like Gap.

Which is, in fact, testing Kindred's robot.

The machine needs to know how to deal

with that new shape.

[George] As winter arrives and we get new objects

we can start learning of those new objects,

and then when summer comes back,

we might see objects we've used before

and we can switch back to those algorithms

to pick up those objects.

Or we might see a whole new class of objects.

I don't know, maybe

sombreros become popular one summer

and now we need to learn how to pick up sombreros.

[Host] Adaptability is crucial for e-commerce robots.

Or any robot for that matter.

We can't just program the machines to manipulate

each and every one of the dizzying number of objects

in our world.

They'll have to think on their own.

And when that doesn't work,

we have to be prepared to step in and help

until they get the hang of things.

Yes, that'll make us babysitters.

But better to babysit than let the machines

get carried away with things.