How to Get Robots to Walk (and Jump and Run) Among Us

[Narrator] Drum roll please.

Well done Cassie.

The AMBER advanced mobility lab at Caltech,

has gotten Cassie the research robot to jump,

which for humans like you and me may seem laughably simple.

But not for robots,

this is an important part of a quest to make robots walk,

bounce and most dramatically jump.

All to unravel the extreme complexities

of bipedal locomotion

because well, humanoid robots have historically had

and 80 year old drunkard kind of vibe.

It turns out that in this lab,

to learn to walk, first you gotta stumble.

Walking on two legs is nothing short of magical.

Cassie walking through fire at the University of Michigan

is even more magical.

Or riding a Segway, but jumping is a whole other challenge.

[Aaron] You have to crouch down,

you have to compress all those springs,

you have to jump off,

you have this air time where you

can't interact with the world at all

and you have to land and then stick that landing.

[Narrator] It's fun to watch, sure,

but it's also a step toward getting Cassie to one day run.

Jumping is a lot like running, in that,

you have a ground phase in which you actuate on the world

and you have a flight phase

where you have to plan your landing

but you can't actually interact with the world.

We wanna make Cassie not just run but run really fast.

[Narrator] That's extremely difficult for a bipedal.

[Aaron] It's this incredibly complex behavior,

you're falling forward and catching yourself continuously.

What's nice about legged robots, especially bipeds,

is they fall pretty dramatically.

So success is a binary you either get good walking

like we do all the time, seemingly simply,

or we fall on our face.

[Narrator] Quadrupeds like MIT's cheetah

are inherently more stable

but bipeds like the famous Atlas from Boston Dynamics

are making rapid progress

and not falling on their faces all the time.

And even one legged robots like

Salto are starting to leap around.

They key here is people like me assaulting them.

We call it disturbance testing,

so let's be clear, it's not assaulting them.

We're testing them with disturbances.

It's actually an important function though,

the real world is full of disturbances

meaning the ground isn't flat, it's rough,

there's weird things that come at you.

So how do you test your algorithms in the way

that they'll translate from the lab environment

to a real world environment.

[Narrator] You mess with your robots is how.

[Aaron] All robots are governed

by the same basic mathematical forms.

What we do is when we develop new techniques

before we bring them to to complex robot

we start with the simpler robots,

we test them out.

[Narrator] So it begins as research here

and ends up here, in Cassie.

But what this lab is learning about

locomotion won't just be reserved

for the speed biped robots of tomorrow.

This is AMPRO a custom design prosthetic leg like no other.

The motors that drive the knee

and the ankle are paired with springs.

AMPRO uses motion sensors to detect where in a gait

the user is, thus syncing human and machine.

[Aaron] So what we do is we generate

walking gaits for this as if it was a robot.

So we start in simulation as we always do,

where we have the model of the human

and the model of the device as they interface together

and then we generate walking gaits for that combined system.

[Narrator] The result is a far more natural stride

than a typical prosthesis.

[Aaron] Piece by pieces we achieve

a new behavior on a robot,

we wanna translate that behavior

over to the prosthetic device to

improve the mobility of the user.

[Narrator] So what begins as a trip

or a bounce, or a leap,

ends up in a new kind of robotic limb.

Almost makes me feel better about what I've done.

Almost.