1
00:00:11,307 --> 00:00:14,491
So what does it mean for a machine to be athletic?

2
00:00:14,491 --> 00:00:17,908
We will demonstrate the concept of machine athleticism

3
00:00:17,908 --> 00:00:19,628
and the research to achieve it

4
00:00:19,628 --> 00:00:22,139
with the help of these flying machines called quadrocopters,

5
00:00:22,139 --> 00:00:24,026
or quads, for short.

6
00:00:26,034 --> 00:00:28,563
Quads have been around for a long time.

7
00:00:28,563 --> 00:00:30,162
The reason that they're so popular these days

8
00:00:30,162 --> 00:00:32,127
is because they're mechanically simple.

9
00:00:32,127 --> 00:00:34,235
By controlling the speeds of these four propellers,

10
00:00:34,235 --> 00:00:37,283
these machines can roll, pitch, yaw,

11
00:00:37,283 --> 00:00:40,108
and accelerate along their common orientation.

12
00:00:40,108 --> 00:00:43,179
On board are also a battery, a computer,

13
00:00:43,179 --> 00:00:47,133
various sensors and wireless radios.

14
00:00:47,133 --> 00:00:51,762
Quads are extremely agile, but this agility comes at a cost.

15
00:00:51,762 --> 00:00:54,754
They are inherently unstable, and they need some form

16
00:00:54,754 --> 00:00:58,726
of automatic feedback control in order to be able to fly.

17
00:01:04,259 --> 00:01:07,497
So, how did it just do that?

18
00:01:07,497 --> 00:01:09,315
Cameras on the ceiling and a laptop

19
00:01:09,315 --> 00:01:11,891
serve as an indoor global positioning system.

20
00:01:11,891 --> 00:01:14,076
It's used to locate objects in the space

21
00:01:14,076 --> 00:01:16,603
that have these reflective markers on them.

22
00:01:16,603 --> 00:01:18,664
This data is then sent to another laptop

23
00:01:18,664 --> 00:01:20,875
that is running estimation and control algorithms,

24
00:01:20,875 --> 00:01:22,761
which in turn sends commands to the quad,

25
00:01:22,761 --> 00:01:26,246
which is also running estimation and control algorithms.

26
00:01:29,620 --> 00:01:32,130
The bulk of our research is algorithms.

27
00:01:32,130 --> 00:01:36,291
It's the magic that brings these machines to life.

28
00:01:36,291 --> 00:01:38,334
So how does one design the algorithms

29
00:01:38,334 --> 00:01:40,611
that create a machine athlete?

30
00:01:40,611 --> 00:01:43,491
We use something broadly called model-based design.

31
00:01:43,491 --> 00:01:46,645
We first capture the physics with a mathematical model

32
00:01:46,645 --> 00:01:48,779
of how the machines behave.

33
00:01:48,779 --> 00:01:51,076
We then use a branch of mathematics

34
00:01:51,076 --> 00:01:53,907
called control theory to analyze these models

35
00:01:53,907 --> 00:01:57,820
and also to synthesize algorithms for controlling them.

36
00:01:57,820 --> 00:02:00,868
For example, that's how we can make the quad hover.

37
00:02:00,868 --> 00:02:02,276
We first captured the dynamics

38
00:02:02,276 --> 00:02:04,182
with a set of differential equations.

39
00:02:04,182 --> 00:02:06,531
We then manipulate these equations with the help

40
00:02:06,531 --> 00:02:11,171
of control theory to create algorithms that stabilize the quad.

41
00:02:11,171 --> 00:02:14,117
Let me demonstrate the strength of this approach.

42
00:02:17,198 --> 00:02:19,747
Suppose that we want this quad to not only hover

43
00:02:19,747 --> 00:02:22,529
but to also balance this pole.

44
00:02:22,529 --> 00:02:24,373
With a little bit of practice,

45
00:02:24,373 --> 00:02:26,764
it's pretty straightforward for a human being to do this,

46
00:02:26,764 --> 00:02:28,519
although we do have the advantage of having

47
00:02:28,519 --> 00:02:29,847
two feet on the ground

48
00:02:29,847 --> 00:02:32,804
and the use of our very versatile hands.

49
00:02:32,804 --> 00:02:35,273
It becomes a little bit more difficult

50
00:02:35,273 --> 00:02:37,569
when I only have one foot on the ground

51
00:02:37,569 --> 00:02:40,368
and when I don't use my hands.

52
00:02:40,368 --> 00:02:43,319
Notice how this pole has a reflective marker on top,

53
00:02:43,319 --> 00:02:47,213
which means that it can be located in the space.

54
00:02:52,971 --> 00:02:58,947
(Applause)

55
00:02:58,947 --> 00:03:01,773
You can notice that this quad is making fine adjustments

56
00:03:01,773 --> 00:03:03,968
to keep the pole balanced.

57
00:03:03,968 --> 00:03:07,027
How did we design the algorithms to do this?

58
00:03:07,027 --> 00:03:09,132
We added the mathematical model of the pole

59
00:03:09,132 --> 00:03:10,651
to that of the quad.

60
00:03:10,651 --> 00:03:13,605
Once we have a model of the combined quad-pole system,

61
00:03:13,605 --> 00:03:18,538
we can use control theory to create algorithms for controlling it.

62
00:03:18,538 --> 00:03:20,126
Here, you see that it's stable,

63
00:03:20,126 --> 00:03:22,819
and even if I give it little nudges,

64
00:03:22,819 --> 00:03:28,222
it goes back to the nice, balanced position.

65
00:03:28,222 --> 00:03:29,885
We can also augment the model to include

66
00:03:29,885 --> 00:03:32,326
where we want the quad to be in space.

67
00:03:32,326 --> 00:03:35,395
Using this pointer, made out of reflective markers,

68
00:03:35,395 --> 00:03:37,619
I can point to where I want the quad to be in space

69
00:03:37,619 --> 00:03:40,986
a fixed distance away from me.

70
00:03:55,763 --> 00:03:58,665
The key to these acrobatic maneuvers is algorithms,

71
00:03:58,665 --> 00:04:01,091
designed with the help of mathematical models

72
00:04:01,091 --> 00:04:03,041
and control theory.

73
00:04:03,041 --> 00:04:05,255
Let's tell the quad to come back here

74
00:04:05,255 --> 00:04:07,074
and let the pole drop,

75
00:04:07,074 --> 00:04:08,992
and I will next demonstrate the importance

76
00:04:08,992 --> 00:04:11,111
of understanding physical models

77
00:04:11,111 --> 00:04:15,231
and the workings of the physical world.

78
00:04:25,462 --> 00:04:27,232
Notice how the quad lost altitude

79
00:04:27,232 --> 00:04:29,314
when I put this glass of water on it.

80
00:04:29,314 --> 00:04:32,115
Unlike the balancing pole, I did not include

81
00:04:32,115 --> 00:04:34,699
the mathematical model of the glass in the system.

82
00:04:34,699 --> 00:04:37,618
In fact, the system doesn't even know that the glass of water is there.

83
00:04:37,618 --> 00:04:40,910
Like before, I could use the pointer to tell the quad

84
00:04:40,910 --> 00:04:43,499
where I want it to be in space.

85
00:04:43,499 --> 00:04:53,096
(Applause)

86
00:04:53,096 --> 00:04:55,206
Okay, you should be asking yourself,

87
00:04:55,206 --> 00:04:57,715
why doesn't the water fall out of the glass?

88
00:04:57,715 --> 00:05:00,926
Two facts: The first is that gravity acts

89
00:05:00,926 --> 00:05:03,051
on all objects in the same way.

90
00:05:03,051 --> 00:05:05,515
The second is that the propellers are all pointing

91
00:05:05,515 --> 00:05:08,642
in the same direction of the glass, pointing up.

92
00:05:08,642 --> 00:05:11,019
You put these two things together, the net result

93
00:05:11,019 --> 00:05:13,306
is that all side forces on the glass are small

94
00:05:13,306 --> 00:05:16,091
and are mainly dominated by aerodynamic effects,

95
00:05:16,091 --> 00:05:19,932
which as these speeds are negligible.

96
00:05:23,336 --> 00:05:25,407
And that's why you don't need to model the glass.

97
00:05:25,407 --> 00:05:29,255
It naturally doesn't spill no matter what the quad does.

98
00:05:38,651 --> 00:05:45,702
(Applause)

99
00:05:45,702 --> 00:05:49,718
The lesson here is that some high-performance tasks

100
00:05:49,718 --> 00:05:51,115
are easier than others,

101
00:05:51,115 --> 00:05:53,411
and that understanding the physics of the problem

102
00:05:53,411 --> 00:05:56,027
tells you which ones are easy and which ones are hard.

103
00:05:56,027 --> 00:05:58,386
In this instance, carrying a glass of water is easy.

104
00:05:58,386 --> 00:06:02,171
Balancing a pole is hard.

105
00:06:02,171 --> 00:06:03,973
We've all heard stories of athletes

106
00:06:03,973 --> 00:06:06,466
performing feats while physically injured.

107
00:06:06,466 --> 00:06:07,985
Can a machine also perform

108
00:06:07,985 --> 00:06:10,636
with extreme physical damage?

109
00:06:10,636 --> 00:06:12,362
Conventional wisdom says that you need

110
00:06:12,362 --> 00:06:16,052
at least four fixed motor propeller pairs in order to fly,

111
00:06:16,052 --> 00:06:18,164
because there are four degrees of freedom to control:

112
00:06:18,164 --> 00:06:21,192
roll, pitch, yaw and acceleration.

113
00:06:21,192 --> 00:06:24,437
Hexacopters and octocopters, with six and eight propellers,

114
00:06:24,437 --> 00:06:25,920
can provide redundancy,

115
00:06:25,920 --> 00:06:27,909
but quadrocopters are much more popular

116
00:06:27,909 --> 00:06:29,724
because they have the minimum number

117
00:06:29,724 --> 00:06:32,200
of fixed motor propeller pairs: four.

118
00:06:32,200 --> 00:06:34,165
Or do they?

119
00:06:49,476 --> 00:06:52,076
If we analyze the mathematical model of this machine

120
00:06:52,076 --> 00:06:54,252
with only two working propellers,

121
00:06:54,252 --> 00:07:01,008
we discover that there's an unconventional way to fly it.

122
00:07:07,980 --> 00:07:09,686
We relinquish control of yaw,

123
00:07:09,686 --> 00:07:12,756
but roll, pitch and acceleration can still be controlled

124
00:07:12,756 --> 00:07:18,208
with algorithms that exploit this new configuration.

125
00:07:21,668 --> 00:07:24,100
Mathematical models tell us exactly when

126
00:07:24,100 --> 00:07:26,172
and why this is possible.

127
00:07:26,172 --> 00:07:28,540
In this instance, this knowledge allows us to design

128
00:07:28,540 --> 00:07:30,796
novel machine architectures

129
00:07:30,796 --> 00:07:34,972
or to design clever algorithms that gracefully handle damage,

130
00:07:34,972 --> 00:07:36,696
just like human athletes do,

131
00:07:36,696 --> 00:07:40,519
instead of building machines with redundancy.

132
00:07:40,519 --> 00:07:42,631
We can't help but hold our breath when we watch

133
00:07:42,631 --> 00:07:45,068
a diver somersaulting into the water,

134
00:07:45,068 --> 00:07:46,858
or when a vaulter is twisting in the air,

135
00:07:46,858 --> 00:07:48,576
the ground fast approaching.

136
00:07:48,576 --> 00:07:51,144
Will the diver be able to pull off a rip entry?

137
00:07:51,144 --> 00:07:53,144
Will the vaulter stick the landing?

138
00:07:53,144 --> 00:07:54,812
Suppose we want this quad here

139
00:07:54,812 --> 00:07:57,263
to perform a triple flip and finish off

140
00:07:57,263 --> 00:07:59,532
at the exact same spot that it started.

141
00:07:59,532 --> 00:08:01,879
This maneuver is going to happen so quickly

142
00:08:01,879 --> 00:08:05,508
that we can't use position feedback to correct the motion during execution.

143
00:08:05,508 --> 00:08:07,782
There simply isn't enough time.

144
00:08:07,782 --> 00:08:11,476
Instead, what the quad can do is perform the maneuver blindly,

145
00:08:11,476 --> 00:08:13,708
observe how it finishes the maneuver,

146
00:08:13,708 --> 00:08:16,132
and then use that information to modify its behavior

147
00:08:16,132 --> 00:08:18,421
so that the next flip is better.

148
00:08:18,421 --> 00:08:20,233
Similar to the diver and the vaulter,

149
00:08:20,233 --> 00:08:22,152
it is only through repeated practice

150
00:08:22,152 --> 00:08:24,077
that the maneuver can be learned and executed

151
00:08:24,077 --> 00:08:26,210
to the highest standard.

152
00:08:34,412 --> 00:08:39,164
(Applause)

153
00:08:39,164 --> 00:08:42,636
Striking a moving ball is a necessary skill in many sports.

154
00:08:42,636 --> 00:08:44,431
How do we make a machine do

155
00:08:44,431 --> 00:08:48,184
what an athlete does seemingly without effort?

156
00:09:03,943 --> 00:09:10,542
(Applause)

157
00:09:10,542 --> 00:09:13,192
This quad has a racket strapped onto its head

158
00:09:13,192 --> 00:09:16,679
with a sweet spot roughly the size of an apple, so not too large.

159
00:09:16,679 --> 00:09:19,846
The following calculations are made every 20 milliseconds,

160
00:09:19,846 --> 00:09:21,602
or 50 times per second.

161
00:09:21,602 --> 00:09:24,414
We first figure out where the ball is going.

162
00:09:24,414 --> 00:09:26,851
We then next calculate how the quad should hit the ball

163
00:09:26,851 --> 00:09:29,950
so that it flies to where it was thrown from.

164
00:09:29,950 --> 00:09:34,406
Third, a trajectory is planned that carries the quad

165
00:09:34,406 --> 00:09:37,430
from its current state to the impact point with the ball.

166
00:09:37,430 --> 00:09:41,239
Fourth, we only execute 20 milliseconds' worth of that strategy.

167
00:09:41,239 --> 00:09:44,046
Twenty milliseconds later, the whole process is repeated

168
00:09:44,046 --> 00:09:46,289
until the quad strikes the ball.

169
00:09:55,560 --> 00:09:58,206
(Applause)

170
00:09:58,206 --> 00:10:01,578
Machines can not only perform dynamic maneuvers on their own,

171
00:10:01,578 --> 00:10:03,441
they can do it collectively.

172
00:10:03,441 --> 00:10:07,028
These three quads are cooperatively carrying a sky net.

173
00:10:16,574 --> 00:10:21,985
(Applause)

174
00:10:21,985 --> 00:10:24,182
They perform an extremely dynamic

175
00:10:24,182 --> 00:10:26,350
and collective maneuver

176
00:10:26,350 --> 00:10:28,042
to launch the ball back to me.

177
00:10:28,042 --> 00:10:31,553
Notice that, at full extension, these quads are vertical.

178
00:10:36,486 --> 00:10:38,258
(Applause)

179
00:10:38,258 --> 00:10:40,545
In fact, when fully extended,

180
00:10:40,545 --> 00:10:43,230
this is roughly five times greater than what a bungee jumper feels

181
00:10:43,230 --> 00:10:47,818
at the end of their launch.

182
00:10:51,009 --> 00:10:53,686
The algorithms to do this are very similar

183
00:10:53,686 --> 00:10:57,070
to what the single quad used to hit the ball back to me.

184
00:10:57,070 --> 00:10:59,934
Mathematical models are used to continuously re-plan

185
00:10:59,934 --> 00:11:04,449
a cooperative strategy 50 times per second.

186
00:11:04,449 --> 00:11:06,326
Everything we have seen so far has been

187
00:11:06,326 --> 00:11:08,789
about the machines and their capabilities.

188
00:11:08,789 --> 00:11:11,804
What happens when we couple this machine athleticism

189
00:11:11,804 --> 00:11:13,580
with that of a human being?

190
00:11:13,580 --> 00:11:17,385
What I have in front of me is a commercial gesture sensor

191
00:11:17,385 --> 00:11:18,694
mainly used in gaming.

192
00:11:18,694 --> 00:11:20,498
It can recognize what my various body parts

193
00:11:20,498 --> 00:11:22,546
are doing in real time.

194
00:11:22,546 --> 00:11:24,662
Similar to the pointer that I used earlier,

195
00:11:24,662 --> 00:11:27,454
we can use this as inputs to the system.

196
00:11:27,454 --> 00:11:29,850
We now have a natural way of interacting

197
00:11:29,850 --> 00:11:34,794
with the raw athleticism of these quads with my gestures.

198
00:12:10,453 --> 00:12:14,853
(Applause)

199
00:12:23,982 --> 00:12:27,989
Interaction doesn't have to be virtual. It can be physical.

200
00:12:27,989 --> 00:12:29,714
Take this quad, for example.

201
00:12:29,714 --> 00:12:32,342
It's trying to stay at a fixed point in space.

202
00:12:32,342 --> 00:12:36,179
If I try to move it out of the way, it fights me,

203
00:12:36,179 --> 00:12:40,334
and moves back to where it wants to be.

204
00:12:40,334 --> 00:12:43,373
We can change this behavior, however.

205
00:12:43,373 --> 00:12:45,072
We can use mathematical models

206
00:12:45,072 --> 00:12:48,205
to estimate the force that I'm applying to the quad.

207
00:12:48,205 --> 00:12:51,443
Once we know this force, we can also change the laws of physics,

208
00:12:51,443 --> 00:12:55,942
as far as the quad is concerned, of course.

209
00:12:55,942 --> 00:12:58,205
Here the quad is behaving as if it were

210
00:12:58,205 --> 00:13:02,570
in a viscous fluid.

211
00:13:02,570 --> 00:13:04,744
We now have an intimate way

212
00:13:04,744 --> 00:13:06,593
of interacting with a machine.

213
00:13:06,593 --> 00:13:09,099
I will use this new capability to position

214
00:13:09,099 --> 00:13:11,676
this camera-carrying quad to the appropriate location

215
00:13:11,676 --> 00:13:14,862
for filming the remainder of this demonstration.

216
00:13:24,222 --> 00:13:26,990
So we can physically interact with these quads

217
00:13:26,990 --> 00:13:29,393
and we can change the laws of physics.

218
00:13:29,393 --> 00:13:31,706
Let's have a little bit of fun with this.

219
00:13:31,706 --> 00:13:33,410
For what you will see next, these quads

220
00:13:33,410 --> 00:13:36,906
will initially behave as if they were on Pluto.

221
00:13:36,906 --> 00:13:39,218
As time goes on, gravity will be increased

222
00:13:39,218 --> 00:13:41,338
until we're all back on planet Earth,

223
00:13:41,338 --> 00:13:43,057
but I assure you we won't get there.

224
00:13:43,057 --> 00:13:46,558
Okay, here goes.

225
00:13:53,667 --> 00:13:57,079
(Laughter)

226
00:14:22,778 --> 00:14:25,986
(Laughter)

227
00:14:25,986 --> 00:14:29,372
(Applause)

228
00:14:29,372 --> 00:14:30,579
Whew!

229
00:14:34,779 --> 00:14:36,296
You're all thinking now,

230
00:14:36,296 --> 00:14:38,271
these guys are having way too much fun,

231
00:14:38,271 --> 00:14:40,350
and you're probably also asking yourself,

232
00:14:40,350 --> 00:14:44,286
why exactly are they building machine athletes?

233
00:14:44,286 --> 00:14:46,938
Some conjecture that the role of play in the animal kingdom

234
00:14:46,938 --> 00:14:49,690
is to hone skills and develop capabilities.

235
00:14:49,690 --> 00:14:51,756
Others think that it has more of a social role,

236
00:14:51,756 --> 00:14:53,430
that it's used to bind the group.

237
00:14:53,430 --> 00:14:56,963
Similarly, we use the analogy of sports and athleticism

238
00:14:56,963 --> 00:14:59,042
to create new algorithms for machines

239
00:14:59,042 --> 00:15:01,189
to push them to their limits.

240
00:15:01,189 --> 00:15:04,866
What impact will the speed of machines have on our way of life?

241
00:15:04,866 --> 00:15:07,255
Like all our past creations and innovations,

242
00:15:07,255 --> 00:15:10,082
they may be used to improve the human condition

243
00:15:10,082 --> 00:15:12,583
or they may be misused and abused.

244
00:15:12,583 --> 00:15:14,506
This is not a technical choice we are faced with;

245
00:15:14,506 --> 00:15:16,261
it's a social one.

246
00:15:16,261 --> 00:15:17,651
Let's make the right choice,

247
00:15:17,651 --> 00:15:20,056
the choice that brings out the best in the future of machines,

248
00:15:20,056 --> 00:15:21,868
just like athleticism in sports

249
00:15:21,868 --> 00:15:24,029
can bring out the best in us.

250
00:15:24,029 --> 00:15:27,479
Let me introduce you to the wizards behind the green curtain.

251
00:15:27,479 --> 00:15:30,379
They're the current members of the Flying Machine Arena research team.

252
00:15:30,379 --> 00:15:35,148
(Applause)

253
00:15:35,148 --> 00:15:38,194
Federico Augugliaro, Dario Brescianini, Markus Hehn,

254
00:15:38,194 --> 00:15:41,018
Sergei Lupashin, Mark Muller and Robin Ritz.

255
00:15:41,018 --> 00:15:42,950
Look out for them. They're destined for great things.

256
00:15:42,950 --> 00:15:44,128
Thank you.

257
00:15:44,128 --> 00:15:50,482
(Applause)