Chris Urmson: How a driverless car sees the road

862,221 views ・ 2015-06-26

TED


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

00:12
So in 1885, Karl Benz invented the automobile.
0
12528
3949
00:16
Later that year, he took it out for the first public test drive,
1
16707
3762
00:20
and -- true story -- crashed into a wall.
2
20469
3375
00:24
For the last 130 years,
3
24184
2043
00:26
we've been working around that least reliable part of the car, the driver.
4
26227
4319
00:30
We've made the car stronger.
5
30546
1354
00:32
We've added seat belts, we've added air bags,
6
32200
2548
00:34
and in the last decade, we've actually started trying to make the car smarter
7
34748
3971
00:38
to fix that bug, the driver.
8
38719
2938
00:41
Now, today I'm going to talk to you a little bit about the difference
9
41657
3261
00:44
between patching around the problem with driver assistance systems
10
44918
3808
00:48
and actually having fully self-driving cars
11
48726
2564
00:51
and what they can do for the world.
12
51290
1880
00:53
I'm also going to talk to you a little bit about our car
13
53170
2995
00:56
and allow you to see how it sees the world and how it reacts and what it does,
14
56165
3999
01:00
but first I'm going to talk a little bit about the problem.
15
60164
3187
01:03
And it's a big problem:
16
63651
1648
01:05
1.2 million people are killed on the world's roads every year.
17
65299
3089
01:08
In America alone, 33,000 people are killed each year.
18
68388
3784
01:12
To put that in perspective,
19
72172
2028
01:14
that's the same as a 737 falling out of the sky every working day.
20
74200
4797
01:19
It's kind of unbelievable.
21
79342
1786
01:21
Cars are sold to us like this,
22
81548
2298
01:23
but really, this is what driving's like.
23
83846
2717
01:26
Right? It's not sunny, it's rainy,
24
86563
2159
01:28
and you want to do anything other than drive.
25
88722
2488
01:31
And the reason why is this:
26
91210
1622
01:32
Traffic is getting worse.
27
92832
1858
01:34
In America, between 1990 and 2010,
28
94690
3506
01:38
the vehicle miles traveled increased by 38 percent.
29
98196
3504
01:42
We grew by six percent of roads,
30
102213
2749
01:44
so it's not in your brains.
31
104962
1602
01:46
Traffic really is substantially worse than it was not very long ago.
32
106564
4276
01:50
And all of this has a very human cost.
33
110840
2409
01:53
So if you take the average commute time in America, which is about 50 minutes,
34
113529
3948
01:57
you multiply that by the 120 million workers we have,
35
117477
3649
02:01
that turns out to be about six billion minutes
36
121126
2225
02:03
wasted in commuting every day.
37
123351
2026
02:05
Now, that's a big number, so let's put it in perspective.
38
125377
2827
02:08
You take that six billion minutes
39
128204
1774
02:09
and you divide it by the average life expectancy of a person,
40
129978
3784
02:13
that turns out to be 162 lifetimes
41
133762
3135
02:16
spent every day, wasted,
42
136897
2925
02:19
just getting from A to B.
43
139822
2044
02:21
It's unbelievable.
44
141866
1730
02:23
And then, there are those of us who don't have the privilege
45
143596
2844
02:26
of sitting in traffic.
46
146440
1672
02:28
So this is Steve.
47
148112
1578
02:29
He's an incredibly capable guy,
48
149690
1765
02:31
but he just happens to be blind,
49
151455
2516
02:33
and that means instead of a 30-minute drive to work in the morning,
50
153971
3217
02:37
it's a two-hour ordeal of piecing together bits of public transit
51
157188
3979
02:41
or asking friends and family for a ride.
52
161167
2385
02:43
He doesn't have that same freedom that you and I have to get around.
53
163552
3669
02:47
We should do something about that.
54
167221
2460
02:49
Now, conventional wisdom would say
55
169891
1757
02:51
that we'll just take these driver assistance systems
56
171648
2492
02:54
and we'll kind of push them and incrementally improve them,
57
174140
3750
02:57
and over time, they'll turn into self-driving cars.
58
177890
2542
03:00
Well, I'm here to tell you that's like me saying
59
180432
2409
03:02
that if I work really hard at jumping, one day I'll be able to fly.
60
182841
4057
03:06
We actually need to do something a little different.
61
186898
2728
03:09
And so I'm going to talk to you about three different ways
62
189626
2711
03:12
that self-driving systems are different than driver assistance systems.
63
192337
3346
03:15
And I'm going to start with some of our own experience.
64
195683
2651
03:18
So back in 2013,
65
198334
2253
03:20
we had the first test of a self-driving car
66
200587
2663
03:23
where we let regular people use it.
67
203250
2027
03:25
Well, almost regular -- they were 100 Googlers,
68
205277
2202
03:27
but they weren't working on the project.
69
207479
2003
03:29
And we gave them the car and we allowed them to use it in their daily lives.
70
209482
3621
03:33
But unlike a real self-driving car, this one had a big asterisk with it:
71
213103
3719
03:36
They had to pay attention,
72
216822
1504
03:38
because this was an experimental vehicle.
73
218326
2633
03:40
We tested it a lot, but it could still fail.
74
220959
3525
03:44
And so we gave them two hours of training,
75
224484
2059
03:46
we put them in the car, we let them use it,
76
226543
2092
03:48
and what we heard back was something awesome,
77
228635
2127
03:50
as someone trying to bring a product into the world.
78
230762
2524
03:53
Every one of them told us they loved it.
79
233286
1925
03:55
In fact, we had a Porsche driver who came in and told us on the first day,
80
235211
3566
03:58
"This is completely stupid. What are we thinking?"
81
238777
2663
04:01
But at the end of it, he said, "Not only should I have it,
82
241850
2840
04:04
everyone else should have it, because people are terrible drivers."
83
244690
3175
04:09
So this was music to our ears,
84
249135
1735
04:10
but then we started to look at what the people inside the car were doing,
85
250870
3803
04:14
and this was eye-opening.
86
254673
1579
04:16
Now, my favorite story is this gentleman
87
256252
2438
04:18
who looks down at his phone and realizes the battery is low,
88
258690
3829
04:22
so he turns around like this in the car and digs around in his backpack,
89
262519
4548
04:27
pulls out his laptop,
90
267067
2153
04:29
puts it on the seat,
91
269220
1567
04:30
goes in the back again,
92
270787
1764
04:32
digs around, pulls out the charging cable for his phone,
93
272551
3367
04:35
futzes around, puts it into the laptop, puts it on the phone.
94
275918
3367
04:39
Sure enough, the phone is charging.
95
279285
2043
04:41
All the time he's been doing 65 miles per hour down the freeway.
96
281328
3994
04:45
Right? Unbelievable.
97
285322
2484
04:47
So we thought about this and we said, it's kind of obvious, right?
98
287806
3121
04:50
The better the technology gets,
99
290927
2263
04:53
the less reliable the driver is going to get.
100
293190
2121
04:55
So by just making the cars incrementally smarter,
101
295311
2396
04:57
we're probably not going to see the wins we really need.
102
297707
2902
05:00
Let me talk about something a little technical for a moment here.
103
300609
3901
05:04
So we're looking at this graph, and along the bottom
104
304510
2438
05:06
is how often does the car apply the brakes when it shouldn't.
105
306948
3051
05:09
You can ignore most of that axis,
106
309999
1621
05:11
because if you're driving around town, and the car starts stopping randomly,
107
311620
3719
05:15
you're never going to buy that car.
108
315339
1701
05:17
And the vertical axis is how often the car is going to apply the brakes
109
317040
3375
05:20
when it's supposed to to help you avoid an accident.
110
320415
3049
05:23
Now, if we look at the bottom left corner here,
111
323464
2221
05:25
this is your classic car.
112
325685
1845
05:27
It doesn't apply the brakes for you, it doesn't do anything goofy,
113
327530
3133
05:30
but it also doesn't get you out of an accident.
114
330663
2779
05:33
Now, if we want to bring a driver assistance system into a car,
115
333442
3018
05:36
say with collision mitigation braking,
116
336460
1828
05:38
we're going to put some package of technology on there,
117
338288
2612
05:40
and that's this curve, and it's going to have some operating properties,
118
340900
3418
05:44
but it's never going to avoid all of the accidents,
119
344318
2490
05:46
because it doesn't have that capability.
120
346808
2059
05:48
But we'll pick some place along the curve here,
121
348867
2249
05:51
and maybe it avoids half of accidents that the human driver misses,
122
351116
3254
05:54
and that's amazing, right?
123
354370
1297
05:55
We just reduced accidents on our roads by a factor of two.
124
355667
2727
05:58
There are now 17,000 less people dying every year in America.
125
358394
3987
06:02
But if we want a self-driving car,
126
362381
2020
06:04
we need a technology curve that looks like this.
127
364401
2307
06:06
We're going to have to put more sensors in the vehicle,
128
366708
2599
06:09
and we'll pick some operating point up here
129
369307
2021
06:11
where it basically never gets into a crash.
130
371328
2019
06:13
They'll happen, but very low frequency.
131
373347
2443
06:15
Now you and I could look at this and we could argue
132
375790
2461
06:18
about whether it's incremental, and I could say something like "80-20 rule,"
133
378251
3605
06:21
and it's really hard to move up to that new curve.
134
381856
2568
06:24
But let's look at it from a different direction for a moment.
135
384424
2934
06:27
So let's look at how often the technology has to do the right thing.
136
387358
3512
06:30
And so this green dot up here is a driver assistance system.
137
390870
3506
06:34
It turns out that human drivers
138
394376
2485
06:36
make mistakes that lead to traffic accidents
139
396861
2647
06:39
about once every 100,000 miles in America.
140
399508
3172
06:42
In contrast, a self-driving system is probably making decisions
141
402680
3167
06:45
about 10 times per second,
142
405847
3663
06:49
so order of magnitude,
143
409510
1422
06:50
that's about 1,000 times per mile.
144
410932
2832
06:53
So if you compare the distance between these two,
145
413764
2485
06:56
it's about 10 to the eighth, right?
146
416249
2600
06:58
Eight orders of magnitude.
147
418849
1765
07:00
That's like comparing how fast I run
148
420614
2809
07:03
to the speed of light.
149
423423
2206
07:05
It doesn't matter how hard I train, I'm never actually going to get there.
150
425629
3785
07:09
So there's a pretty big gap there.
151
429414
2438
07:11
And then finally, there's how the system can handle uncertainty.
152
431852
3729
07:15
So this pedestrian here might be stepping into the road, might not be.
153
435581
3323
07:18
I can't tell, nor can any of our algorithms,
154
438904
3395
07:22
but in the case of a driver assistance system,
155
442310
2284
07:24
that means it can't take action, because again,
156
444594
2806
07:27
if it presses the brakes unexpectedly, that's completely unacceptable.
157
447400
3339
07:30
Whereas a self-driving system can look at that pedestrian and say,
158
450739
3133
07:33
I don't know what they're about to do,
159
453872
1890
07:35
slow down, take a better look, and then react appropriately after that.
160
455762
3762
07:39
So it can be much safer than a driver assistance system can ever be.
161
459524
3702
07:43
So that's enough about the differences between the two.
162
463226
2730
07:45
Let's spend some time talking about how the car sees the world.
163
465956
3484
07:49
So this is our vehicle.
164
469440
1252
07:50
It starts by understanding where it is in the world,
165
470692
2438
07:53
by taking a map and its sensor data and aligning the two,
166
473130
2787
07:55
and then we layer on top of that what it sees in the moment.
167
475917
2948
07:58
So here, all the purple boxes you can see are other vehicles on the road,
168
478865
3655
08:02
and the red thing on the side over there is a cyclist,
169
482520
2528
08:05
and up in the distance, if you look really closely,
170
485048
2402
08:07
you can see some cones.
171
487450
1794
08:09
Then we know where the car is in the moment,
172
489244
2773
08:12
but we have to do better than that: we have to predict what's going to happen.
173
492017
3833
08:15
So here the pickup truck in top right is about to make a left lane change
174
495850
3488
08:19
because the road in front of it is closed,
175
499338
2223
08:21
so it needs to get out of the way.
176
501561
1731
08:23
Knowing that one pickup truck is great,
177
503292
1863
08:25
but we really need to know what everybody's thinking,
178
505155
2479
08:27
so it becomes quite a complicated problem.
179
507634
2507
08:30
And then given that, we can figure out how the car should respond in the moment,
180
510141
4749
08:34
so what trajectory it should follow, how quickly it should slow down or speed up.
181
514890
3866
08:38
And then that all turns into just following a path:
182
518756
3065
08:41
turning the steering wheel left or right, pressing the brake or gas.
183
521821
3197
08:45
It's really just two numbers at the end of the day.
184
525018
2464
08:47
So how hard can it really be?
185
527482
2241
08:50
Back when we started in 2009,
186
530433
1952
08:52
this is what our system looked like.
187
532385
1798
08:54
So you can see our car in the middle and the other boxes on the road,
188
534183
3391
08:57
driving down the highway.
189
537574
1271
08:58
The car needs to understand where it is and roughly where the other vehicles are.
190
538845
3818
09:02
It's really a geometric understanding of the world.
191
542663
2429
09:05
Once we started driving on neighborhood and city streets,
192
545092
2948
09:08
the problem becomes a whole new level of difficulty.
193
548040
2445
09:10
You see pedestrians crossing in front of us, cars crossing in front of us,
194
550485
3494
09:13
going every which way,
195
553979
1811
09:15
the traffic lights, crosswalks.
196
555790
1527
09:17
It's an incredibly complicated problem by comparison.
197
557317
2797
09:20
And then once you have that problem solved,
198
560114
2103
09:22
the vehicle has to be able to deal with construction.
199
562217
2512
09:24
So here are the cones on the left forcing it to drive to the right,
200
564729
3151
09:27
but not just construction in isolation, of course.
201
567880
2402
09:30
It has to deal with other people moving through that construction zone as well.
202
570282
3723
09:34
And of course, if anyone's breaking the rules, the police are there
203
574005
3263
09:37
and the car has to understand that that flashing light on the top of the car
204
577268
3622
09:40
means that it's not just a car, it's actually a police officer.
205
580890
3105
09:43
Similarly, the orange box on the side here,
206
583995
2032
09:46
it's a school bus,
207
586027
1109
09:47
and we have to treat that differently as well.
208
587136
2520
09:50
When we're out on the road, other people have expectations:
209
590576
2793
09:53
So, when a cyclist puts up their arm,
210
593369
1780
09:55
it means they're expecting the car to yield to them and make room for them
211
595149
3518
09:58
to make a lane change.
212
598667
2053
10:01
And when a police officer stood in the road,
213
601030
2173
10:03
our vehicle should understand that this means stop,
214
603203
2740
10:05
and when they signal to go, we should continue.
215
605943
3506
10:09
Now, the way we accomplish this is by sharing data between the vehicles.
216
609449
3761
10:13
The first, most crude model of this
217
613210
1696
10:14
is when one vehicle sees a construction zone,
218
614906
2113
10:17
having another know about it so it can be in the correct lane
219
617019
3062
10:20
to avoid some of the difficulty.
220
620081
1570
10:21
But we actually have a much deeper understanding of this.
221
621651
2664
10:24
We could take all of the data that the cars have seen over time,
222
624315
3009
10:27
the hundreds of thousands of pedestrians, cyclists,
223
627324
2376
10:29
and vehicles that have been out there
224
629700
1787
10:31
and understand what they look like
225
631487
1695
10:33
and use that to infer what other vehicles should look like
226
633182
2831
10:36
and other pedestrians should look like.
227
636013
1926
10:37
And then, even more importantly, we could take from that a model
228
637939
3021
10:40
of how we expect them to move through the world.
229
640960
2330
10:43
So here the yellow box is a pedestrian crossing in front of us.
230
643290
2963
10:46
Here the blue box is a cyclist and we anticipate
231
646253
2250
10:48
that they're going to nudge out and around the car to the right.
232
648503
3312
10:52
Here there's a cyclist coming down the road
233
652115
2092
10:54
and we know they're going to continue to drive down the shape of the road.
234
654207
3486
10:57
Here somebody makes a right turn,
235
657693
1867
10:59
and in a moment here, somebody's going to make a U-turn in front of us,
236
659560
3360
11:02
and we can anticipate that behavior and respond safely.
237
662920
2614
11:05
Now, that's all well and good for things that we've seen,
238
665534
2728
11:08
but of course, you encounter lots of things that you haven't
239
668262
2865
11:11
seen in the world before.
240
671127
1231
11:12
And so just a couple of months ago,
241
672358
1741
11:14
our vehicles were driving through Mountain View,
242
674099
2235
11:16
and this is what we encountered.
243
676334
1644
11:17
This is a woman in an electric wheelchair
244
677978
2082
11:20
chasing a duck in circles on the road. (Laughter)
245
680060
2617
11:22
Now it turns out, there is nowhere in the DMV handbook
246
682677
3111
11:25
that tells you how to deal with that,
247
685788
2245
11:28
but our vehicles were able to encounter that,
248
688033
2143
11:30
slow down, and drive safely.
249
690176
2255
11:32
Now, we don't have to deal with just ducks.
250
692431
2041
11:34
Watch this bird fly across in front of us. The car reacts to that.
251
694472
3708
11:38
Here we're dealing with a cyclist
252
698180
1615
11:39
that you would never expect to see anywhere other than Mountain View.
253
699795
3290
11:43
And of course, we have to deal with drivers,
254
703085
2068
11:45
even the very small ones.
255
705153
3715
11:48
Watch to the right as someone jumps out of this truck at us.
256
708868
4131
11:54
And now, watch the left as the car with the green box decides
257
714460
2929
11:57
he needs to make a right turn at the last possible moment.
258
717389
3325
12:00
Here, as we make a lane change, the car to our left decides
259
720714
2851
12:03
it wants to as well.
260
723565
3553
12:07
And here, we watch a car blow through a red light
261
727118
2693
12:09
and yield to it.
262
729811
2090
12:11
And similarly, here, a cyclist blowing through that light as well.
263
731901
3854
12:15
And of course, the vehicle responds safely.
264
735755
2746
12:18
And of course, we have people who do I don't know what
265
738501
2601
12:21
sometimes on the road, like this guy pulling out between two self-driving cars.
266
741102
3823
12:24
You have to ask, "What are you thinking?"
267
744925
2045
12:26
(Laughter)
268
746970
1212
12:28
Now, I just fire-hosed you with a lot of stuff there,
269
748182
2521
12:30
so I'm going to break one of these down pretty quickly.
270
750703
2650
12:33
So what we're looking at is the scene with the cyclist again,
271
753353
2940
12:36
and you might notice in the bottom, we can't actually see the cyclist yet,
272
756293
3491
12:39
but the car can: it's that little blue box up there,
273
759784
2504
12:42
and that comes from the laser data.
274
762288
2081
12:44
And that's not actually really easy to understand,
275
764369
2418
12:46
so what I'm going to do is I'm going to turn that laser data and look at it,
276
766787
3584
12:50
and if you're really good at looking at laser data, you can see
277
770371
3029
12:53
a few dots on the curve there,
278
773400
1487
12:54
right there, and that blue box is that cyclist.
279
774887
2372
12:57
Now as our light is red,
280
777259
1149
12:58
the cyclist's light has turned yellow already,
281
778408
2192
13:00
and if you squint, you can see that in the imagery.
282
780600
2438
13:03
But the cyclist, we see, is going to proceed through the intersection.
283
783038
3286
13:06
Our light has now turned green, his is solidly red,
284
786324
2394
13:08
and we now anticipate that this bike is going to come all the way across.
285
788718
4292
13:13
Unfortunately the other drivers next to us were not paying as much attention.
286
793010
3742
13:16
They started to pull forward, and fortunately for everyone,
287
796752
3157
13:19
this cyclists reacts, avoids,
288
799909
3011
13:22
and makes it through the intersection.
289
802920
2191
13:25
And off we go.
290
805111
1568
13:26
Now, as you can see, we've made some pretty exciting progress,
291
806679
2948
13:29
and at this point we're pretty convinced
292
809627
1902
13:31
this technology is going to come to market.
293
811529
2010
13:33
We do three million miles of testing in our simulators every single day,
294
813539
4783
13:38
so you can imagine the experience that our vehicles have.
295
818322
2689
13:41
We are looking forward to having this technology on the road,
296
821011
2864
13:43
and we think the right path is to go through the self-driving
297
823875
2890
13:46
rather than driver assistance approach
298
826765
1844
13:48
because the urgency is so large.
299
828609
2621
13:51
In the time I have given this talk today,
300
831230
2393
13:53
34 people have died on America's roads.
301
833623
3135
13:56
How soon can we bring it out?
302
836758
2368
13:59
Well, it's hard to say because it's a really complicated problem,
303
839126
3832
14:02
but these are my two boys.
304
842958
2214
14:05
My oldest son is 11, and that means in four and a half years,
305
845172
3623
14:08
he's going to be able to get his driver's license.
306
848795
2577
14:11
My team and I are committed to making sure that doesn't happen.
307
851372
3204
14:14
Thank you.
308
854576
1904
14:16
(Laughter) (Applause)
309
856480
3667
14:21
Chris Anderson: Chris, I've got a question for you.
310
861110
2568
14:23
Chris Urmson: Sure.
311
863678
2809
14:26
CA: So certainly, the mind of your cars is pretty mind-boggling.
312
866487
3924
14:30
On this debate between driver-assisted and fully driverless --
313
870411
4459
14:34
I mean, there's a real debate going on out there right now.
314
874870
3041
14:37
So some of the companies, for example, Tesla,
315
877911
2833
14:40
are going the driver-assisted route.
316
880744
2159
14:42
What you're saying is that that's kind of going to be a dead end
317
882903
5248
14:48
because you can't just keep improving that route and get to fully driverless
318
888151
5456
14:53
at some point, and then a driver is going to say, "This feels safe,"
319
893607
3530
14:57
and climb into the back, and something ugly will happen.
320
897137
2647
14:59
CU: Right. No, that's exactly right, and it's not to say
321
899784
2676
15:02
that the driver assistance systems aren't going to be incredibly valuable.
322
902460
3537
15:05
They can save a lot of lives in the interim,
323
905997
2058
15:08
but to see the transformative opportunity to help someone like Steve get around,
324
908055
3833
15:11
to really get to the end case in safety,
325
911888
1969
15:13
to have the opportunity to change our cities
326
913857
2479
15:16
and move parking out and get rid of these urban craters we call parking lots,
327
916336
4204
15:20
it's the only way to go.
328
920540
1240
15:21
CA: We will be tracking your progress with huge interest.
329
921780
2718
15:24
Thanks so much, Chris. CU: Thank you. (Applause)
330
924498
4232
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