1
00:00:00,000 --> 00:00:02,530


2
00:00:02,530 --> 00:00:05,350
This is Toddler.

3
00:00:05,350 --> 00:00:07,990
This is Dr. Russ Tedrake.

4
00:00:07,990 --> 00:00:11,200
Together, they're covering
new ground at MIT.

5
00:00:11,200 --> 00:00:14,470
So we call the robot Toddler
because it learns to walk

6
00:00:14,470 --> 00:00:16,990
and because it
toddles when it walks.

7
00:00:16,990 --> 00:00:18,820
These are baby steps
towards developing

8
00:00:18,820 --> 00:00:22,600
a new class of robot, one
that uses very little energy,

9
00:00:22,600 --> 00:00:26,200
about the same as a
student walking to class.

10
00:00:26,200 --> 00:00:29,440
There are no complicated gears
and hydraulics for Toddler.

11
00:00:29,440 --> 00:00:32,650
It's a simple mechanical
device, actually based

12
00:00:32,650 --> 00:00:35,110
on an old design
for walking toys.

13
00:00:35,110 --> 00:00:37,510
They're called passive
dynamic walkers.

14
00:00:37,510 --> 00:00:41,590
They waddled down hills, using
gravity as their power source.

15
00:00:41,590 --> 00:00:43,630
Toddler can do
that, but its makers

16
00:00:43,630 --> 00:00:45,790
have added a little extra power.

17
00:00:45,790 --> 00:00:48,370
It has a computer and
battery pack on board,

18
00:00:48,370 --> 00:00:51,430
but it does a lot of the
work itself, gaining momentum

19
00:00:51,430 --> 00:00:53,500
from its own steps.

20
00:00:53,500 --> 00:00:56,050
And Toddler doesn't
just walk downhill,

21
00:00:56,050 --> 00:00:59,140
he can walk on
flat surfaces, too.

22
00:00:59,140 --> 00:01:01,180
We're letting the
dynamics of the robot

23
00:01:01,180 --> 00:01:02,830
solve most of the
control problem,

24
00:01:02,830 --> 00:01:04,413
and we just add a
little bit of energy

25
00:01:04,413 --> 00:01:06,580
here there to kind of
push and pull the robot,

26
00:01:06,580 --> 00:01:09,490
and that's a much more
energy efficient strategy

27
00:01:09,490 --> 00:01:11,500
for designing a walking robot.

28
00:01:11,500 --> 00:01:14,170
That little bit of energy
makes a big difference.

29
00:01:14,170 --> 00:01:16,660
High-performance
robots like ASIMO

30
00:01:16,660 --> 00:01:18,740
may be capable of many tasks.

31
00:01:18,740 --> 00:01:20,740
But they're loaded
with control programs,

32
00:01:20,740 --> 00:01:22,710
and those demand a lot of juice.

33
00:01:22,710 --> 00:01:26,830
ASIMO's batteries need to be
recharged every half hour,

34
00:01:26,830 --> 00:01:28,330
but not Toddler.

35
00:01:28,330 --> 00:01:32,440
It's just such a thrill whenever
you see the robot walking,

36
00:01:32,440 --> 00:01:34,870
and especially when you do it
in, like, in a public place

37
00:01:34,870 --> 00:01:36,662
and you can see all
the people go, oh look,

38
00:01:36,662 --> 00:01:38,920
it's the cool little robot.

39
00:01:38,920 --> 00:01:40,600
But what makes Toddler
really intriguing

40
00:01:40,600 --> 00:01:44,380
is that like a human youngster,
this robot can learn.

41
00:01:44,380 --> 00:01:47,680
Toddler's sensors communicate
directly with its motors.

42
00:01:47,680 --> 00:01:49,780
They sense a change in
the walking surface,

43
00:01:49,780 --> 00:01:53,205
allowing the robot to make
adjustments with every step.

44
00:01:53,205 --> 00:01:54,580
And that's what
we call learning.

45
00:01:54,580 --> 00:01:57,220
It's a simple idea,
but it's the best way

46
00:01:57,220 --> 00:02:00,070
we can embody what we think the
nervous system is doing, too,

47
00:02:00,070 --> 00:02:01,960
when we're talking
about learning.

48
00:02:01,960 --> 00:02:04,990
Researchers can adjust the
pace at which Toddler learns,

49
00:02:04,990 --> 00:02:07,420
but faster learning
comes with a price.

50
00:02:07,420 --> 00:02:09,503
Toddler can move from
linoleum to carpet,

51
00:02:09,503 --> 00:02:11,920
quickly learning the difference
and adjusting its movement

52
00:02:11,920 --> 00:02:13,000
and balance.

53
00:02:13,000 --> 00:02:15,310
But if it has to move
back to linoleum,

54
00:02:15,310 --> 00:02:17,590
takes the robot a little longer.

55
00:02:17,590 --> 00:02:19,180
There's this very
interesting result

56
00:02:19,180 --> 00:02:20,860
in learning theory,
which tells you,

57
00:02:20,860 --> 00:02:23,470
you can't have learning
without generalization.

58
00:02:23,470 --> 00:02:25,618
The problem is when
you start generalizing,

59
00:02:25,618 --> 00:02:27,910
then there's a possibility
that you'll overwrite things

60
00:02:27,910 --> 00:02:31,020
that you've learned in the past.

61
00:02:31,020 --> 00:02:32,760
Toddler is one of
three new walking

62
00:02:32,760 --> 00:02:34,290
robots that are
testing the limits

63
00:02:34,290 --> 00:02:36,720
of passive dynamic design.

64
00:02:36,720 --> 00:02:39,840
This one at Cornell
University, in New York state,

65
00:02:39,840 --> 00:02:41,850
is testing efficiency.

66
00:02:41,850 --> 00:02:45,510
The design allows this robot to
push off with its trailing leg,

67
00:02:45,510 --> 00:02:48,270
giving it a little extra boost.

68
00:02:48,270 --> 00:02:50,850
And this robot at Delft
University in the Netherlands

69
00:02:50,850 --> 00:02:53,400
is testing robustness of design.

70
00:02:53,400 --> 00:02:55,440
It's able to maintain
a stable gait,

71
00:02:55,440 --> 00:02:57,568
even when pushed off balance.

72
00:02:57,568 --> 00:02:59,610
This robot works on the
principle of the best way

73
00:02:59,610 --> 00:03:01,890
to stop yourself
from falling forward

74
00:03:01,890 --> 00:03:03,810
is to put your foot out.

75
00:03:03,810 --> 00:03:05,730
The recovery motion
is made possible

76
00:03:05,730 --> 00:03:09,780
by a hip actuator that
acts like a hip joint.

77
00:03:09,780 --> 00:03:12,150
These three teams are hoping
to meld their research

78
00:03:12,150 --> 00:03:15,660
to create one robot that
is efficient, robust,

79
00:03:15,660 --> 00:03:17,460
and capable of learning.

80
00:03:17,460 --> 00:03:21,330
It's our goal to convince people
of the value of this approach

81
00:03:21,330 --> 00:03:23,940
and to just keep building
better and better robots to do

82
00:03:23,940 --> 00:03:28,110
more and more things, and
all exploit the dynamics.

83
00:03:28,110 --> 00:03:29,580
So how does studying
a robot that

84
00:03:29,580 --> 00:03:32,760
walks like a human
actually benefit humans?

85
00:03:32,760 --> 00:03:35,640
Ultimately, down the
line, far down the line,

86
00:03:35,640 --> 00:03:38,340
if we understand the motor
learning system better, then

87
00:03:38,340 --> 00:03:40,590
there's a potential that we
could use that information

88
00:03:40,590 --> 00:03:43,440
to design rehabilitation
programs, even

89
00:03:43,440 --> 00:03:46,620
neural prosthetics, to help
people recover from injuries

90
00:03:46,620 --> 00:03:48,640
or augment their performance.

91
00:03:48,640 --> 00:03:50,730
There's a lot of different
potential applications.

92
00:03:50,730 --> 00:03:52,770
As for Toddler, he's got
a big brother already

93
00:03:52,770 --> 00:03:55,200
in the works, one
with smaller feet

94
00:03:55,200 --> 00:03:58,740
and a fine set of knees
that can lock and swing.

95
00:03:58,740 --> 00:04:01,380
And that will make the
robot more versatile

96
00:04:01,380 --> 00:04:03,450
and also, just a little
more human, something

97
00:04:03,450 --> 00:04:05,760
Dr. Tedrake thinks is important.

98
00:04:05,760 --> 00:04:08,010
When you see a robot
that looks like a human,

99
00:04:08,010 --> 00:04:09,720
you have a tendency
to reach out to it.

100
00:04:09,720 --> 00:04:11,580
And it really
helps, I think, with

101
00:04:11,580 --> 00:04:13,530
the human-robot interaction.

102
00:04:13,530 --> 00:04:14,940
[MUSIC PLAYING]

103
00:04:14,940 --> 00:04:17,820
I think it's the first step
in a long road of building

104
00:04:17,820 --> 00:04:19,560
better robots, more
efficient robots,

105
00:04:19,560 --> 00:04:21,060
and learning systems for robots.

106
00:04:21,060 --> 00:04:24,390
So I hope that it's just
the first piece of a very

107
00:04:24,390 --> 00:04:27,290
complex and very long puzzle.

108
00:04:27,290 --> 00:04:34,000