"Go even harder!"
"Seriously?" "Yeah!" That's Atlas, the super famous robot
that does parkour and throws stuff in these incredible videos from Boston Dynamics.
And I'm about to shove it as hard as I possibly can. And then, to be fair, I'm
going to get shoved the same way. I'm here with Atlas for a little friendly competition to see
what this robot can really do and to answer a question that I've had since I was a kid: When
are all the robots coming? You know, like C3PO and Sonny from "I, Robot" and The Iron Giant. The human-
looking machines that can help us do more than we can by ourselves... as long as they don't try and
take over. For most of my life, humanoid robots have seemed pretty far away. But recently? Not
so much! Tesla, Boston Dynamics, Honda, NASA, and a bunch of startups are all making huge leaps
in this field. So in this video we're going to take you to the cutting edge of humanoid robots
to show you what they can actually do right now and what you can expect in
the robot future that's coming... "Humanoid robots..." "imitation of life..." "machines that resemble people
as closely as possible" "robots playing a larger role in our lives" "how do we prevent robots and humans
from being in conflict?" "How do I look? Great let's do this." I've wanted to meet Atlas for years. I'm one of
the hundreds of millions of people who have watched this robot on YouTube.
But getting into Boston Dynamics is hard. I've been trying for months and
finally as the last ditch effort I just tweeted asking anyone who watches this
show for help. Through this community, we were able to connect to the right people and finally
Boston Dynamics said yes. So you're the reason I get to be here. Thank you for making
this dream of mine come true. "I'm so excited to be here.
Thank you for taking the time." "We're excited to have you. This is the Atlas lab.
You recognize it from our YouTube videos?" "I do! Very much. Can I meet Atlas for the first time?" "Do you still feel this way
when you stand next to Atlas?" "Every day, I love this robot." The reason that I wanted to meet Atlas so badly is
that this robot represents the much larger effort across many companies to build machines that resemble
us. That walk on two legs, and use arms and hands and can do some of the same things that we can and
maybe even things that we can't. If you understand Atlas, you can understand what's really happening
with humanoid robots as a whole. Let's start with some anatomy. "Head, Shoulders, Knees, and Toes," but
for robots. Atlas's eyes are where you'd expect and this is what the world looks like to it. Atlas's
brain though is not in its head. It's in its chest. "Atlas has three computers inside of its torso. Each
of which is the equivalent of a beefy laptop basically." Also, Atlas was kind of... short. "I didn't realize that I would be taller than Atlas!
But Atlas is BUILT." This version of Atlas is 4' 11 inches or
1.5 M tall. That's about the height of the average American 12-year-old. Early
Boston Dynamics robots were much larger. The larger versions though were just scarier. It's easier
to find a robot endearing when it's the height of a middle schooler. No one's ever seen the most
famous humanoid robots from different companies standing next to each other but based on their
specs this is what that would look like: Here's Atlas. Here's me at 5'9. Tesla's Optimus robot
is 5'8. NASA's Valkyrie robot is bigger 6'2. And the cute old Honda Asimo is little just 4'3. And
there are plenty more. My biggest question to these companies though was: Why? Why build a
human-shaped robot? We already have robots in all kinds of other shapes. Why are we working so hard
to make them vaguely like us? The answer boils down to this: We built our world for humans. "So if you're interested in building
a system that can start to address the current limitations in automation, a
humanoid form factor is quite good." Companies are building these robots to first do things
that are unsafe or repetitive or boring for humans like some kinds of manufacturing or
some tasks in hard to get places. That's why legs specifically are so important... "So many of the robots that people
encounter day to day from a Roomba to a more automated delivery system,
they're on wheels. Why have legs?" "For me, legs are about access. The opportunity to go more places is just better. So
if I want a robot that can not only walk around a floor like this but that can go upstairs, climb
ladders, maybe I can take it into the woods, or I can take it into a crumbling building..." Or in the case of NASA's Valkyrie
robot, they want to take it into space! Talk about uncertain terrain... "Why not add something humans
don't have... like a tail?" "You have to think about what the function of a tail
might be. The tail would have to be able to exert forces on the robot that would be significant..." "I'm just thinking about all the things
that I wanted as a kid...." "I wanted wings also. Is
Atlas going to go airborne anytime?" "That would be pretty sweet, um..." [CRICKETS] "Yeah maybe." Yeah, that's a no. I gotta ask the questions!
It's my job. Ok so it's valuable for robots to have have legs and
a human- shaped body, even if it doesn't come with wings or a tail. But what really matters is
what it can do. Time to put Atlas to the test! But before we get to that let me show you something: So
I've been traveling more for this show especially internationally and when I do that I always want
to make my computer think that I'm still in the US, just so that everything online feels normal. My
websites don't get blocked and things are still in dollars, stuff like that. The way that I do that
is with a virtual private network or "VPN" like Surfshark. Here I'll show you: I can set it to the
US when I'm abroad or I can set it to somewhere else when I'm at home. Look at this. I can set it
to one of over a hundred different places. And the other thing is when I'm using Surfshark, it's
protecting me in the background. So for example it helps prevent unauthorized access to my camera
and notifies me when someone tries accessing it. If you don't use a VPN, you might find it useful.
I do. And right now you can get an exclusive Surfshark Black Friday deal. Enter promo code "Cleo"
to get up to six additional months for free at Surfshark dot deals slash Cleo. Now where were we? Ready? Let's go. First up: running.
I crushed Atlas at running. Jumping? We can both do that... And also this... Or rolling... gotta say,
Atlas is more graceful than I am. And then... The grand finale... Show-off. "Do a backflip!" [ATTEMPT] "Nope! There's no way." "Jesus Christ. That's Jason Borne." Yeah! No! I can't do a standing backflip.
There's just no way. But... hang on... if Atlas can do a
standing backflip... why can't it beat me at something as simple
as running? I was going really slow. "Have you ever heard of Moravec's Paradox?" "No." "It's a principle that's often
quoted in robotics. It's that the easy things are hard. So basically
all the things that we take for granted in how we move through the world world and interact with
objects, we spend no time thinking about, these are exactly the things that are
extremely hard to get to work in robots." When you see a person do something
impressive, you subconsciously make assumptions about what else they can do. If they can do a backflip, they must be able
to jog slowly. The lesson here though is: Don't do that for robots!
Those assumptions don't apply. "So if I see a person for example
take an amazing soccer shot that goes in the top left corner
of the goal and it looks beautiful, I could probably assume that that person is really good
at dribbling soccer balls. But if you see a robot do that those other things might not be true at all!" Atlas can do a backflip
but doesn't sit in a char. "Humanoids don't sit in chairs. And
there's actually technical reasons for it. It's very hard to write a a general control system
that can just sit in a chair. If we spent a lot of time and worked on them, I'm sure we
could do it, but I think there's just a lot of really simple, low performance,
interactive behaviors that we do all day long that would be a real
struggle for Atlas right now." So far though, Atlas and I have
competed in predictable environments. But what about when something unexpected
happens? Like hypothetically some mean YouTuber coming to shove you over? "I feel guilty already. Where do I shove it?" "You can shove it right in
the camera, that's fine." "Stop, really?"
"Yeah, go ahead, it's fine" "Wow!" "I'm shoving it hard too..." "Go even harder!" "Seriously?" "Yeah!" "I can't!" "Wow. I'm actually trying.
Hard. Ok I'm going to to try one more time and then I'll
stop bullying your robot." "Go ahead!" "Wow... I'm so sorry." If the robots do ever take over, we're
just gonna agree that this never happened. Okay? Okay. And besides, in my defense,
I asked one of the engineers to shove me in the same way that I shoved Atlas... "Would you push me in the chest?" "I'm serious. I want to be in
the same exact position that Atlas was... Listen, fair's fair. But not in the head,
that felt like a little much. You can see that I recover too in almost
exactly the same way that Atlas does but I get pushed back way farther so I'd say Atlas wins this
round. These are the moments that really pique my interest. How does a robot handle adversity? Like
in Tesla's recent Optimus demo, one of its blocks fell unexpectedly on its side and it seamlessly
corrected it. And then this guy comes in and moves the poor robots blocks around, and it just keeps
going, it's just adapting to the changes. Also I felt so bad for it. Like don't move the poor robots
blocks, around he's organizing them! And I felt so guilty for pushing Atlas too... "Like I actually feel guilt.
I'm not trying to be cute on camera, I feel something when I push a humanoid robot." "I think there's just going to be a big
learning experience for us as a society as these systems start to make
their way into our daily lives." Obviously I know that Atlas is not a person.
But we're people, so it's just inevitable that we respond differently to robots that look like us as
opposed to say an automated forklift. This is one of the things that I think is most interesting
about this topic. Most of us subconsciously think if they look more like us they must be and think
more like us but how intelligent is Atlas really? when I see Atlas throw a bag on YouTube has a
human engineer pre-planned every tiny little movement of each limb? Or have they said, "get this
bag from here to here, but you figure out how"? Or could they just say, "We're building a house. Help
us do that." And then Atlas would identify what to do? That might be the future, but the answer right
now is closest to that middle option: "Get this from here to here." "The way that we command the robot
to throw the bag is we tell the robot okay we want this bag to end up over there. You get to hold
it for some period of time, and we want you to do a 180. So it has to figure out, okay how do I grab
the bag and then accelerate it while I'm jumping so that when I let it go it follows Newton's laws
and ends up where it's supposed to go. And all that coordination is really complicated right?" For you as a person, this doesn't
require much thinking at all. "And then you're going to jump and turn 180
and throw it behind you." "Okay... everybody... beware. That way." "Wow, that's heavier when I...
let me try again." "Turns out I can't do that as high as I thought." As you noticed, when you did a 180 without the bag you did it perfectly, right? As soon as you
did that your balance is all off and the reason is the inertia of this bag coupled with your body
changes how much angular momentum you need to do in order to get all the way around right and the
robot has to figure out all of those details on the fly." "This is making me more impressed with the
level of computation that Atlas is doing and that my brain is doing. And I'm appreciating my
own body more like a robot right now." Recreating human-like movement in a robot is just
very hard and Atlas doesn't always get this right. Plus you're learning from every new
experience and failure. Atlas is not. "When I'm shoving Atlas as hard as possible,
is Atlas learning to accommodate that better? Or would someone need to change Atlas's
programming in order to accommodate that better?" "Yeah right now we're not automatically using
data from the robot to improve the control system so any improvement that comes from hardware
experiments goes through a human engineer at this point." Okay so Atlas isn't learning from every new
experience. But it is learning in another way... "How much is Atlas currently using
what a layperson would call AI?" "So when people say "AI," they're usually
referring to models that were created through a process of machine learning..." Machine learning boils down to
giving a machine a set of examples of inputs and outputs and allowing it to figure
out its own rules to get from one to the other. "And so Atlas's perception system for example
is almost entirely driven by machine learning at this point. So it frequently uses cameras in order
to identify objects in its environment in order to localize itself to its surroundings. There's
other types of machine learning as well that are more responsible for generating behavior.
Instead of a human engineer painstakingly thinking carefully about you know the physics of
a problem and writing down a controller to solve that problem, you actually use simulation
and maybe robot data in some cases and trial and error experience in order for the system to
learn how to do a task better." Based on its demo videos, Tesla appears to be heavily
investing in exactly that. When they say Optimus can sort these blocks "fully autonomously," what they seem to mean
is that after being given the goal of "blue blocks here and green blocks here" the robot can then use
a neural network, a kind of machine learning system, that runs inside the robot to process
visual information from its cameras and "decide" how to accomplish this task. Tesla also seems to be
using simulation data to train Optimus to perform tasks better. Frankly it's hard from these videos
to figure out exactly how Optimus is working but I'd love to visit Tesla and find out! The robot
future that I wanted as a kid, having robots in our lives helping us do daily tasks, will depend on
these robots being able to understand increasingly high level concepts and commands. If you want a
robot to put away your groceries there is a massive difference between "rotate your left
hand, pick up the bag like this, put the bag on the counter, then take your left hand and open..."
the versus "put my groceries away." A robot needs to understand not just "where are my hands?" but "what
are groceries?" and "away" means "in the cabinet, not out the window" and also other things remain more
important at the same time, like don't put away my groceries at all cost! Don't step on my dog!
The more you learn about humanoid robots the more you realize how difficult making one actually is,
and how amazing the robots that already exist are, and also how amazing human bodies are. But as
these robots inevitably get more capable one question will become more urgent:
What will we use them for? "Killer robots" "Will robots take over the world?" "The global robot takeover is coming!" Boston Dynamics has issued a
statement saying "no general purpose robots should be weaponized." I asked them about that. "One of the things that I'm proud of at this company is
that we've taking a very strong stance for example against weaponization of robots, collaborating with
Massachusetts legislators in order to put forward legislation that would explicitly ban the creation
of such technology in the state." It seems like Elon Musk has a similar stance on this.
He said that we don't want a "terminator scenario" but if history and Hollywood are any indication not all
robot makers might feel this way. This is likely to be an ongoing conversation within these
companies and between all of us. "I think many of us on this team are are looking for
ways to make these have a positive impact in the world. Basically allowing people to decide long-term
what kind of work they want to do. And then for the things that we don't want to do we can have machines
that'll help us keep things going." There are robots around us already, all the time.
Most of them static or on wheels, but humanoid robots are inherently different.
Something about making them look like us fundamentally changes
how we feel about them. They are more scary and more lovable because they're
more like us. Making machines that look like us is one of the most ambitious and endearingly
human technological moonshots I can think of. It's a creation in our own image. There is a "huge if true"
future where humanoid robots are walking among us, helping us by doing things that we do
like taking out groceries or building things and doing things that we can't, like running into
the worst burning building or walking unprotected on the moon. Hundreds of people are are working
right now to make that future possible and if they succeed, we're going to need to think about
another question which is: What do we as humans want to do - and what do we want to give up? We have
some time before we need to answer that question but the robot future is coming and it's up to us
to decide what it looks like. Hold on... I have one more thing... For this whole video we were
talking about robots that look like humans but what about robots that look... like dogs? Oh thank
you! I got to hang out with Boston Dynamics robot dogs and you definitely want to see that video
so subscribe to support our show and to see it first.
No comments yet. Be the first to comment!