TED2007

Hod Lipson: Building "self-aware" robots

Filmed:

Hod Lipson demonstrates a few of his cool little robots, which have the ability to learn, understand themselves and even self-replicate.

- Roboticist
Hod Lipson works at the intersection of engineering and biology, studying robots and the way they "behave" and evolve. His work has exciting implications for design and manufacturing -- and serves as a window to understand our own behavior and evolution. Full bio

So, where are the robots?
00:25
We've been told for 40 years already that they're coming soon.
00:27
Very soon they'll be doing everything for us.
00:30
They'll be cooking, cleaning, buying things, shopping, building. But they aren't here.
00:33
Meanwhile, we have illegal immigrants doing all the work,
00:38
but we don't have any robots.
00:42
So what can we do about that? What can we say?
00:44
So I want to give a little bit of a different perspective
00:48
of how we can perhaps look at these things in a little bit of a different way.
00:52
And this is an x-ray picture
00:58
of a real beetle, and a Swiss watch, back from '88. You look at that --
01:00
what was true then is certainly true today.
01:05
We can still make the pieces. We can make the right pieces.
01:07
We can make the circuitry of the right computational power,
01:10
but we can't actually put them together to make something
01:13
that will actually work and be as adaptive as these systems.
01:16
So let's try to look at it from a different perspective.
01:21
Let's summon the best designer, the mother of all designers.
01:23
Let's see what evolution can do for us.
01:27
So we threw in -- we created a primordial soup
01:30
with lots of pieces of robots -- with bars, with motors, with neurons.
01:34
Put them all together, and put all this under kind of natural selection,
01:38
under mutation, and rewarded things for how well they can move forward.
01:42
A very simple task, and it's interesting to see what kind of things came out of that.
01:46
So if you look, you can see a lot of different machines
01:52
come out of this. They all move around.
01:55
They all crawl in different ways, and you can see on the right,
01:57
that we actually made a couple of these things,
02:01
and they work in reality. These are not very fantastic robots,
02:03
but they evolved to do exactly what we reward them for:
02:06
for moving forward. So that was all done in simulation,
02:10
but we can also do that on a real machine.
02:13
Here's a physical robot that we actually
02:15
have a population of brains,
02:20
competing, or evolving on the machine.
02:23
It's like a rodeo show. They all get a ride on the machine,
02:25
and they get rewarded for how fast or how far
02:28
they can make the machine move forward.
02:31
And you can see these robots are not ready
02:33
to take over the world yet, but
02:35
they gradually learn how to move forward,
02:38
and they do this autonomously.
02:40
So in these two examples, we had basically
02:43
machines that learned how to walk in simulation,
02:47
and also machines that learned how to walk in reality.
02:50
But I want to show you a different approach,
02:52
and this is this robot over here, which has four legs.
02:54
It has eight motors, four on the knees and four on the hip.
03:00
It has also two tilt sensors that tell the machine
03:02
which way it's tilting.
03:05
But this machine doesn't know what it looks like.
03:08
You look at it and you see it has four legs,
03:10
the machine doesn't know if it's a snake, if it's a tree,
03:12
it doesn't have any idea what it looks like,
03:14
but it's going to try to find that out.
03:17
Initially, it does some random motion,
03:19
and then it tries to figure out what it might look like.
03:21
And you're seeing a lot of things passing through its minds,
03:24
a lot of self-models that try to explain the relationship
03:26
between actuation and sensing. It then tries to do
03:30
a second action that creates the most disagreement
03:33
among predictions of these alternative models,
03:37
like a scientist in a lab. Then it does that
03:39
and tries to explain that, and prune out its self-models.
03:41
This is the last cycle, and you can see it's pretty much
03:45
figured out what its self looks like. And once it has a self-model,
03:48
it can use that to derive a pattern of locomotion.
03:52
So what you're seeing here are a couple of machines --
03:56
a pattern of locomotion.
03:58
We were hoping that it wass going to have a kind of evil, spidery walk,
04:00
but instead it created this pretty lame way of moving forward.
04:04
But when you look at that, you have to remember
04:08
that this machine did not do any physical trials on how to move forward,
04:11
nor did it have a model of itself.
04:17
It kind of figured out what it looks like, and how to move forward,
04:19
and then actually tried that out.
04:22
(Applause)
04:26
So, we'll move forward to a different idea.
04:31
So that was what happened when we had a couple of --
04:35
that's what happened when you had a couple of -- OK, OK, OK --
04:40
(Laughter)
04:44
-- they don't like each other. So
04:46
there's a different robot.
04:48
That's what happened when the robots actually
04:51
are rewarded for doing something.
04:53
What happens if you don't reward them for anything, you just throw them in?
04:55
So we have these cubes, like the diagram showed here.
04:58
The cube can swivel, or flip on its side,
05:01
and we just throw 1,000 of these cubes into a soup --
05:04
this is in simulation --and don't reward them for anything,
05:08
we just let them flip. We pump energy into this
05:10
and see what happens in a couple of mutations.
05:13
So, initially nothing happens, they're just flipping around there.
05:16
But after a very short while, you can see these blue things
05:19
on the right there begin to take over.
05:23
They begin to self-replicate. So in absence of any reward,
05:25
the intrinsic reward is self-replication.
05:29
And we've actually built a couple of these,
05:32
and this is part of a larger robot made out of these cubes.
05:33
It's an accelerated view, where you can see the robot actually
05:37
carrying out some of its replication process.
05:40
So you're feeding it with more material -- cubes in this case --
05:42
and more energy, and it can make another robot.
05:46
So of course, this is a very crude machine,
05:49
but we're working on a micro-scale version of these,
05:52
and hopefully the cubes will be like a powder that you pour in.
05:54
OK, so what can we learn? These robots are of course
05:57
not very useful in themselves, but they might teach us something
06:02
about how we can build better robots,
06:05
and perhaps how humans, animals, create self-models and learn.
06:08
And one of the things that I think is important
06:13
is that we have to get away from this idea
06:15
of designing the machines manually,
06:17
but actually let them evolve and learn, like children,
06:19
and perhaps that's the way we'll get there. Thank you.
06:22
(Applause)
06:24

▲Back to top

About the Speaker:

Hod Lipson - Roboticist
Hod Lipson works at the intersection of engineering and biology, studying robots and the way they "behave" and evolve. His work has exciting implications for design and manufacturing -- and serves as a window to understand our own behavior and evolution.

Why you should listen

To say that Hod Lipson and his team at Cornell build robots is not completely accurate: They may simply set out a pile of virtual robot parts, devise some rules for assembly, and see what the parts build themselves into. They've created robots that decide for themselves how they want to walk; robots that develop a sense of what they look like; even robots that can, through trial and error, construct other robots just like themselves.

Working across disciplines -- physics, computer science, math, biology and several flavors of engineer -- the team studies techniques for self-assembly and evolution that have great implications for fields such as micro-manufacturing -- allowing tiny pieces to assemble themselves at scales heretofore impossible -- and extreme custom manufacturing (in other words, 3-D printers for the home).

His lab's Outreach page is a funhouse of tools and instructions, including the amazing Golem@Home -- a self-assembling virtual robot who lives in your screensaver.

More profile about the speaker
Hod Lipson | Speaker | TED.com