How AI Will Step Off the Screen and into the Real World | Daniela Rus | TED

233,197 views ・ 2024-04-19

TED


Please double-click on the English subtitles below to play the video.

00:04
When I was a student studying robotics,
0
4334
2377
00:06
a group of us decided to make a present for our professor's birthday.
1
6711
4046
00:11
We wanted to program our robot to cut a slice of cake for him.
2
11508
4212
00:16
We pulled an all-nighter writing the software,
3
16638
3170
00:19
and the next day, disaster.
4
19849
2253
00:22
We programmed this robot to cut a soft, round sponge cake,
5
22769
4546
00:27
but we didn't coordinate well.
6
27357
1710
00:29
And instead, we received a square hard ice cream cake.
7
29109
4588
00:34
The robot flailed wildly and nearly destroyed the cake.
8
34239
3879
00:38
(Laughter)
9
38159
1377
00:39
Our professor was delighted, anyway.
10
39578
2335
00:41
He calmly pushed the stop button
11
41955
3045
00:45
and declared the erratic behavior of the robot
12
45041
3128
00:48
a control singularity.
13
48169
1836
00:50
A robotics technical term.
14
50046
1710
00:52
I was disappointed, but I learned a very important lesson.
15
52882
3879
00:56
The physical world,
16
56761
1794
00:58
with its physics laws and imprecisions,
17
58555
2502
01:01
is a far more demanding space than the digital world.
18
61057
3462
01:05
Today, I lead MIT's Computer Science and AI lab,
19
65478
3838
01:09
the largest research unit at MIT.
20
69357
2419
01:11
This is our buildingm where I work with brilliant and brave researchers
21
71776
5214
01:16
to invent the future of computing and intelligent machines.
22
76990
3545
01:21
Today in computing,
23
81286
1168
01:22
artificial intelligence and robotics are largely separate fields.
24
82454
4087
01:27
AI has amazed you with its decision-making and learning,
25
87375
4129
01:31
but it remains confined inside computers.
26
91546
2878
01:34
Robots have a physical presence and can execute pre-programmed tasks,
27
94883
4796
01:39
but they're not intelligent.
28
99679
1794
01:42
Well, this separation is starting to change.
29
102140
2669
01:45
AI is about to break free from the 2D computer screen interactions
30
105268
4713
01:50
and enter a vibrant, physical 3D world.
31
110023
3462
01:54
In my lab, we're fusing the digital intelligence of AI
32
114361
3837
01:58
with the mechanical prowess of robots.
33
118198
2377
02:01
Moving AI from the digital world into the physical world
34
121034
2836
02:03
is making machines intelligent
35
123912
2127
02:06
and leading to the next great breakthrough,
36
126081
2419
02:08
what I call physical intelligence.
37
128541
2253
02:11
Physical intelligence is when AI's power to understand text,
38
131586
4505
02:16
images and other online information
39
136132
2419
02:18
is used to make real-world machines smarter.
40
138593
3128
02:21
This means AI can help pre-programmed robots do their tasks better
41
141721
5381
02:27
by using knowledge from data.
42
147143
1877
02:31
With physical intelligence,
43
151022
1460
02:32
AI doesn't just reside in our computers,
44
152482
4713
02:37
but walks, rolls, flies
45
157237
2502
02:39
and interacts with us in surprising ways.
46
159781
2961
02:42
Imagine being surrounded by helpful robots at the supermarket.
47
162784
4630
02:47
The one on the left can help you carry a heavy box.
48
167414
3044
02:51
To make it happen, we need to do a few things.
49
171251
3378
02:54
We need to rethink how machines think.
50
174671
2377
02:57
We need to reorganize how they are designed and how they learn.
51
177382
4796
03:03
So for physical intelligence,
52
183596
1585
03:05
AI has to run on computers that fit on the body of the robot.
53
185223
4129
03:09
For example, our soft robot fish.
54
189853
2502
03:13
Today's AI uses server farms that do not fit.
55
193189
3128
03:17
Today's AI also makes mistakes.
56
197318
3212
03:20
This AI system on a robot car does not detect pedestrians.
57
200572
4171
03:25
For physical intelligence,
58
205660
1418
03:27
we need small brains that do not make mistakes.
59
207120
2961
03:31
We're tackling these challenges using inspiration
60
211958
2836
03:34
from a worm called C. elegans
61
214794
2169
03:37
In sharp contrast to the billions of neurons in the human brain,
62
217839
4630
03:42
C. elegans has a happy life on only 302 neurons,
63
222469
4546
03:47
and biologists understand the math of what each of these neurons do.
64
227015
4171
03:53
So here's the idea.
65
233354
1168
03:54
Can we build AI using inspiration from the math of these neurons?
66
234564
5547
04:01
We have developed, together with my collaborators and students,
67
241529
4255
04:05
a new approach to AI we call “liquid networks.”
68
245784
3670
04:10
And liquid networks results in much more compact
69
250121
3754
04:13
and explainable solutions than today's traditional AI solutions.
70
253917
3962
04:17
Let me show you.
71
257921
1251
04:19
This is our self-driving car.
72
259464
1918
04:21
It's trained using a traditional AI solution,
73
261800
2669
04:24
the kind you find in many applications today.
74
264511
2836
04:28
This is the dashboard of the car.
75
268097
2086
04:30
In the lower right corner, you'll see the map.
76
270225
2294
04:32
In the upper left corner, the camera input stream.
77
272560
3254
04:35
And the big box in the middle with the blinking lights
78
275814
2836
04:38
is the decision-making engine.
79
278691
2169
04:40
It consists of tens of thousands of artificial neurons,
80
280902
4004
04:44
and it decides how the car should steer.
81
284948
2502
04:48
It is impossible to correlate the activity of these neurons
82
288076
3336
04:51
with the behavior of the car.
83
291454
2211
04:53
Moreover, if you look at the lower left side,
84
293706
3379
04:57
you see where in the image this decision-making engine looks
85
297085
4045
05:01
to tell the car what to do.
86
301172
2086
05:03
And you see how noisy it is.
87
303299
1418
05:04
And this car drives by looking at the bushes and the trees
88
304759
4254
05:09
on the side of the road.
89
309013
1460
05:10
That's not how we drive.
90
310473
1418
05:11
People look at the road.
91
311933
1335
05:13
Now contrast this with our liquid network solution,
92
313643
3253
05:16
which consists of only 19 neurons rather than tens of thousands.
93
316938
4922
05:21
And look at its attention map.
94
321860
1543
05:23
It's so clean and focused on the road horizon
95
323403
2752
05:26
and the side of the road.
96
326197
1669
05:28
Because these models are so much smaller,
97
328491
2294
05:30
we actually understand how they make decisions.
98
330827
2669
05:34
So how did we get this performance?
99
334831
2586
05:38
Well, in a traditional AI system,
100
338418
2752
05:41
the computational neuron is the artificial neuron,
101
341170
3003
05:44
and the artificial neuron is essentially an on/off computational unit.
102
344215
4213
05:48
It takes in some numbers, adds them up,
103
348469
2211
05:50
applies some basic math
104
350680
1293
05:52
and passes along the result.
105
352015
2002
05:54
And this is complex
106
354058
1335
05:55
because it happens across thousands of computational units.
107
355435
3712
05:59
In liquid networks,
108
359439
1585
06:01
we have fewer neurons,
109
361065
1377
06:02
but each one does more complex math.
110
362483
2711
06:05
Here's what happens inside our liquid neuron.
111
365194
2628
06:08
We use differential equations to model the neural computation
112
368239
3921
06:12
and the artificial synapse.
113
372201
1669
06:14
And these differential equations
114
374412
2085
06:16
are what biologists have mapped for the neural structure of the worms.
115
376539
5089
06:22
We also wire the neurons differently to increase the information flow.
116
382337
4963
06:27
Well, these changes yield phenomenal results.
117
387675
3045
06:31
Traditional AI systems are frozen after training.
118
391054
3420
06:34
That means they cannot continue to improve
119
394515
2294
06:36
when we deploy them in a physical world in the wild.
120
396809
3379
06:40
We just wait for the next release.
121
400229
2253
06:43
Because of what's happening inside the liquid neuron,
122
403316
3378
06:46
liquid networks continue to adapt after training
123
406736
2920
06:49
based on the inputs that they see.
124
409697
1752
06:51
Let me show you.
125
411449
1293
06:53
We trained traditional AI and liquid networks
126
413493
3086
06:56
using summertime videos like these ones,
127
416621
3253
06:59
and the task was to find things in the woods.
128
419916
3045
07:02
All the models learned how to do the task in the summer.
129
422961
3044
07:06
Then we tried to use the models on drones in the fall.
130
426589
3754
07:10
The traditional AI solution gets confused by the background.
131
430343
3837
07:14
Look at the attention map, cannot do the task.
132
434222
2836
07:17
Liquid networks do not get confused by the background
133
437350
3170
07:20
and very successfully execute the task.
134
440520
4004
07:24
So this is it.
135
444899
1168
07:26
This is the step forward:
136
446109
1334
07:27
AI that adapts after training.
137
447443
2670
07:31
Liquid networks are important
138
451072
2044
07:33
because they give us a new way of getting machines to think
139
453116
5088
07:38
that is rooted into physics models,
140
458246
2669
07:40
a new technology for AI.
141
460957
2044
07:43
We can run them on smartphones, on robots,
142
463418
3003
07:46
on enterprise computers,
143
466462
2169
07:48
and even on new types of machines
144
468631
2252
07:50
that we can now begin to imagine and design.
145
470925
2669
07:53
The second aspect of physical intelligence.
146
473594
2753
07:56
So by now you've probably generated images using text-to-image systems.
147
476848
5589
08:02
We can also do text-to-robot,
148
482437
1918
08:04
but not using today's AI solutions because they work on statistics
149
484397
3962
08:08
and do not understand physics.
150
488359
1960
08:11
In my lab,
151
491154
1167
08:12
we developed an approach that guides the design process
152
492363
4004
08:16
by checking and simulating the physical constraints for the machine.
153
496409
4838
08:21
We start with a language prompt,
154
501706
1877
08:23
"Make me a robot that can walk forward,"
155
503583
2502
08:26
and our system generates the designs including shape, materials, actuators,
156
506085
6090
08:32
sensors, the program to control it
157
512175
3003
08:35
and the fabrication files to make it.
158
515178
2294
08:37
And then the designs get refined in simulation
159
517805
3254
08:41
until they meet the specifications.
160
521100
2753
08:44
So in a few hours we can go from idea
161
524312
3670
08:48
to controllable physical machine.
162
528024
2294
08:51
We can also do image-to-robot.
163
531486
1960
08:53
This photo can be transformed into a cuddly robotic bunny.
164
533488
4629
08:58
To do so, our algorithm computes a 3D representation of the photo
165
538618
5297
09:03
that gets sliced and folded, printed.
166
543915
4254
09:08
Then we fold the printed layers, we string some motors and sensors.
167
548169
4338
09:12
We write some code, and we get the bunny you see in this video.
168
552548
3504
09:16
We can use this approach to make anything almost,
169
556844
3379
09:20
from an image, from a photo.
170
560264
2169
09:23
So the ability to transform text into images
171
563309
4922
09:28
and to transform images into robots is important,
172
568231
3253
09:31
because we are drastically reducing the amount of time
173
571484
3920
09:35
and the resources needed to prototype and test new products,
174
575404
3796
09:39
and this is allowing for a much faster innovation cycle.
175
579200
5255
09:45
And now we are ready to even make the leap
176
585164
3587
09:48
to get these machines to learn.
177
588751
1752
09:50
The third aspect of physical intelligence.
178
590545
3044
09:54
These machines can learn from humans how to do tasks.
179
594507
2753
09:57
You can think of it as human-to-robot.
180
597260
2377
09:59
In my lab, we created a kitchen environment
181
599929
2753
10:02
where we instrument people with sensors,
182
602723
2294
10:05
and we collect a lot of data about how people do kitchen tasks.
183
605017
4213
10:09
We need physical data
184
609689
2043
10:11
because videos do not capture the dynamics of the task.
185
611774
4004
10:15
So we collect muscle, pose, even gaze information
186
615820
3170
10:18
about how people do tasks.
187
618990
2043
10:21
And then we train AI using this data
188
621075
3462
10:24
to teach robots how to do the same tasks.
189
624579
2711
10:28
And the end result is machines that move with grace and agility,
190
628541
5589
10:34
as well as adapt and learn.
191
634172
2335
10:36
Physical intelligence.
192
636549
1668
10:39
We can use this approach to teach robots
193
639177
3044
10:42
how to do a wide range of tasks:
194
642263
2294
10:44
food preparation, cleaning and so much more.
195
644557
3170
10:49
The ability to turn images and text into functional machines,
196
649312
5630
10:54
coupled with using liquid networks
197
654984
1960
10:56
to create powerful brains for these machines
198
656986
2336
10:59
that can learn from humans, is incredibly exciting.
199
659322
3211
11:02
Because this means we can make almost anything we imagine.
200
662533
4505
11:07
Today's AI has a ceiling.
201
667663
2086
11:09
It requires server farms.
202
669790
1377
11:11
It's not sustainable.
203
671167
1293
11:12
It makes inexplicable mistakes.
204
672501
2503
11:15
Let's not settle for the current offering.
205
675046
2460
11:18
When AI moves into the physical world,
206
678132
2461
11:20
the opportunities for benefits and for breakthroughs is extraordinary.
207
680635
4838
11:26
You can get personal assistants that optimize your routines
208
686849
4505
11:31
and anticipate your needs,
209
691354
1835
11:33
bespoke machines that help you at work
210
693606
2919
11:36
and robots that delight you in your spare time.
211
696525
3170
11:40
The promise of physical intelligence is to transcend our human limitations
212
700321
5380
11:45
with capabilities that extend our reach,
213
705743
3045
11:48
amplify our strengths
214
708829
1710
11:50
and refine our precision
215
710539
2128
11:52
and grant us ways to interact with the world
216
712667
3086
11:55
we've only dreamed of.
217
715795
1668
11:58
We are the only species so advanced, so aware,
218
718547
4296
12:02
so capable of building these extraordinary tools.
219
722843
3295
12:06
Yet, developing physical intelligence
220
726973
2919
12:09
is teaching us that we have so much more to learn
221
729934
2544
12:12
about technology and about ourselves.
222
732520
3003
12:16
We need human guiding hands over AI sooner rather than later.
223
736232
4588
12:20
After all, we remain responsible for this planet
224
740820
3086
12:23
and everything living on it.
225
743948
1918
12:26
I remain convinced that we have the power
226
746450
3128
12:29
to use physical intelligence to ensure a better future for humanity
227
749620
5339
12:34
and for the planet.
228
754959
1460
12:36
And I'd like to invite you to help us in this quest.
229
756460
3462
12:39
Some of you will help develop physical intelligence.
230
759964
3337
12:43
Some of you will use it.
231
763301
2002
12:45
And some of you will invent the future.
232
765344
2795
12:48
Thank you.
233
768139
1168
12:49
(Applause)
234
769348
4672
About this website

This site will introduce you to YouTube videos that are useful for learning English. You will see English lessons taught by top-notch teachers from around the world. Double-click on the English subtitles displayed on each video page to play the video from there. The subtitles scroll in sync with the video playback. If you have any comments or requests, please contact us using this contact form.

https://forms.gle/WvT1wiN1qDtmnspy7