1
00:00:01,333 --> 00:00:03,133
NARRATOR:
A lethal technology

2
00:00:03,166 --> 00:00:04,600
we can't live without.

3
00:00:04,633 --> 00:00:06,833
(tires screeching,
horns honking)

4
00:00:08,600 --> 00:00:11,266
But what if cars got smarter?

5
00:00:11,300 --> 00:00:14,733
CHRIS GERDES:
Automated vehicles offer
the promise

6
00:00:14,766 --> 00:00:17,966
of dramatically reducing
collisions and fatalities

7
00:00:18,000 --> 00:00:20,133
on our roads and highways.

8
00:00:20,166 --> 00:00:21,300
The big dream is to make a car

9
00:00:21,333 --> 00:00:23,966
that will never be responsible
for a collision.

10
00:00:24,000 --> 00:00:27,366
NARRATOR:
The potential payoff is huge.

11
00:00:27,400 --> 00:00:30,700
RAJ RAJKUMAR:
In the mid-2030s,
the market could be worth

12
00:00:30,733 --> 00:00:33,566
a few trillion U.S. dollars,
with a T in there.

13
00:00:33,600 --> 00:00:35,733
NARRATOR:
Pursuing that pot of gold,

14
00:00:35,766 --> 00:00:40,733
companies are already testing
their cars on our streets.

15
00:00:40,766 --> 00:00:43,766
This is a bold
and ambitious mission.

16
00:00:43,800 --> 00:00:46,200
NARRATOR:
Is this the beginning
of a mobility revolution?

17
00:00:46,233 --> 00:00:47,766
AMNON SHASHUA:
We're venturing into domains

18
00:00:47,800 --> 00:00:50,400
that ten years ago
would be deemed science fiction.

19
00:00:50,433 --> 00:00:54,733
NARRATOR:
Or are we simply moving
too fast?

20
00:00:54,766 --> 00:00:57,033
MISSY CUMMINGS:
The technologies
are deeply flawed.

21
00:00:57,066 --> 00:01:01,200
It's just simply not ready
for public consumption.

22
00:01:01,233 --> 00:01:07,333
NARRATOR:
Will we ever be safe
with a robot behind the wheel?

23
00:01:07,366 --> 00:01:10,633
"Look Who's Driving,"

24
00:01:10,666 --> 00:01:12,466
right now, on "NOVA."

25
00:01:12,500 --> 00:01:16,666
♪

26
00:01:18,700 --> 00:01:21,066
Major funding for "NOVA"
is provided by the following:

27
00:01:37,400 --> 00:01:40,200
♪

28
00:01:40,233 --> 00:01:43,333
OPERATOR (on phone):
911, what is your emergency?

29
00:01:43,366 --> 00:01:45,066
RAFAELA VASQUEZ (on phone):
Um, yes, I, um...

30
00:01:45,100 --> 00:01:46,533
I hit a bicyclist.

31
00:01:46,566 --> 00:01:49,000
OPERATOR:
Do you need paramedics?

32
00:01:49,033 --> 00:01:51,000
VASQUEZ:
I don't, but they do.

33
00:01:51,033 --> 00:01:53,066
♪

34
00:01:53,100 --> 00:01:55,100
NARRATOR:
Every day in the United States,

35
00:01:55,133 --> 00:01:59,066
there are about 100
fatal car crashes.

36
00:01:59,100 --> 00:02:02,000
But on March 18, 2018,

37
00:02:02,033 --> 00:02:04,733
one attracts
particular attention,

38
00:02:04,766 --> 00:02:08,766
because the woman behind
the wheel, Rafaela Vasquez,

39
00:02:08,800 --> 00:02:11,100
is not driving--

40
00:02:11,133 --> 00:02:13,600
a computer is.

41
00:02:13,633 --> 00:02:15,433
♪

42
00:02:19,000 --> 00:02:23,533
The crash puts a spotlight on
a controversial new technology:

43
00:02:23,566 --> 00:02:25,633
the self-driving car.

44
00:02:25,666 --> 00:02:27,533
POLICE OFFICER:
So what exactly happened?

45
00:02:27,566 --> 00:02:29,366
VASQUEZ:
Well, the car
was in auto-drive.

46
00:02:29,400 --> 00:02:30,333
POLICE OFFICER:
Uh-huh.

47
00:02:30,366 --> 00:02:31,566
VASQUEZ:
And all of a sudden, I...

48
00:02:31,600 --> 00:02:33,533
The car didn't see it,
I didn't see it.

49
00:02:33,566 --> 00:02:35,000
And all of a sudden,
she's just there.

50
00:02:35,033 --> 00:02:37,033
But she shot out in front.

51
00:02:37,066 --> 00:02:38,833
And then I think I...

52
00:02:38,866 --> 00:02:40,233
I know I hit 'em.

53
00:02:40,266 --> 00:02:45,433
NARRATOR:
The victim is 49-year-old
Elaine Herzberg.

54
00:02:45,466 --> 00:02:48,400
The car that kills her
is being tested by Uber,

55
00:02:48,433 --> 00:02:51,366
the ride-hailing company.

56
00:02:51,400 --> 00:02:54,600
It's a modified Volvo equipped

57
00:02:54,633 --> 00:02:57,933
with Uber's self-driving
technology.

58
00:02:57,966 --> 00:02:59,900
Test drives are permitted
in Arizona,

59
00:02:59,933 --> 00:03:02,600
as long as there's
a safety driver to take over

60
00:03:02,633 --> 00:03:04,700
in case of trouble.
(brakes squeak)

61
00:03:04,733 --> 00:03:06,433
♪

62
00:03:06,466 --> 00:03:11,166
But on this night,
Uber's experiment badly fails.

63
00:03:11,200 --> 00:03:15,866
Elaine Herzberg is
the first person in history

64
00:03:15,900 --> 00:03:20,000
killed by a self-driving car.

65
00:03:20,033 --> 00:03:25,666
♪

66
00:03:27,600 --> 00:03:31,166
A woman was pushing a bike
across a road.

67
00:03:31,200 --> 00:03:33,966
This is the sweet spot,
in theory,

68
00:03:34,000 --> 00:03:35,966
where autonomy would be
at its best,

69
00:03:36,000 --> 00:03:37,166
particularly compared to humans,

70
00:03:37,200 --> 00:03:39,133
who have terrible vision
at night.

71
00:03:39,166 --> 00:03:43,633
And I think that's a, sadly, um,
it's a very good example

72
00:03:43,666 --> 00:03:48,500
of just how far away from safe
these cars really are.

73
00:03:48,533 --> 00:03:50,466
POLICE OFFICER 2:
Now, the question I have
for you guys is,

74
00:03:50,500 --> 00:03:52,366
do you guys have remote access

75
00:03:52,400 --> 00:03:54,033
to the cameras and stuff
like that that are in there?

76
00:03:54,066 --> 00:03:55,900
UBER REP:
We technically should, yeah.

77
00:03:55,933 --> 00:03:59,600
NARRATOR:
Herzberg's death quickly becomes
an international story.

78
00:03:59,633 --> 00:04:02,633
REPORTER:
Killed on the street
by a self-driving Uber car...

79
00:04:02,666 --> 00:04:04,866
REPORTER:
Uber is now suspending all
of its self-driving testing...

80
00:04:04,900 --> 00:04:07,533
REPORTER:
Self-driving cars
under intense scrutiny after...

81
00:04:07,566 --> 00:04:10,533
Uber's C.E.O. tweeting,
"Incredibly sad news.

82
00:04:10,566 --> 00:04:12,466
We're thinking of
the victim's family..."

83
00:04:12,500 --> 00:04:16,333
REPORTER:
So the question here is,
"What went wrong?"

84
00:04:16,366 --> 00:04:18,566
♪

85
00:04:18,600 --> 00:04:20,433
NARRATOR:
The Tempe crash stokes
public fears

86
00:04:20,466 --> 00:04:22,366
about self-driving cars.

87
00:04:22,400 --> 00:04:26,366
Nearly three out of four
Americans

88
00:04:26,400 --> 00:04:28,100
say they'd be afraid
to ride in one.

89
00:04:28,133 --> 00:04:30,800
So why is anyone even trying

90
00:04:30,833 --> 00:04:34,533
to make a car
that drives itself?

91
00:04:34,566 --> 00:04:37,800
Self-driving vehicles
don't get distracted.

92
00:04:37,833 --> 00:04:41,300
They don't get fatigued,
uh, they don't fall asleep.

93
00:04:41,333 --> 00:04:43,533
Uh, and, you know,
they don't drive drunk.

94
00:04:43,566 --> 00:04:48,300
NARRATOR:
In other words,
they promise safety.

95
00:04:48,333 --> 00:04:50,266
Each year in the U.S.,

96
00:04:50,300 --> 00:04:55,100
some 35,000 people die
in car crashes,

97
00:04:55,133 --> 00:04:58,500
nearly all caused
by human error.

98
00:04:58,533 --> 00:05:00,433
They're because of a choice
or an error that we make,

99
00:05:00,466 --> 00:05:01,900
a choice to pick up the phone,

100
00:05:01,933 --> 00:05:05,100
drive drunk, drive drugged,
drive distracted, drive drowsy.

101
00:05:05,133 --> 00:05:07,266
(horn honking)

102
00:05:07,300 --> 00:05:08,633
You're looking one direction

103
00:05:08,666 --> 00:05:10,666
when you should have been
watching in the other direction.

104
00:05:10,700 --> 00:05:12,200
94% of crashes.

105
00:05:13,533 --> 00:05:16,133
NARRATOR:
The hope is that computers

106
00:05:16,166 --> 00:05:18,233
will be able to do a better job.

107
00:05:18,266 --> 00:05:19,733
DANIELA RUS:
I really believe

108
00:05:19,766 --> 00:05:21,333
that with self-driving cars,

109
00:05:21,366 --> 00:05:23,866
we will eliminate
road accidents.

110
00:05:23,900 --> 00:05:26,300
If we get the technology right,

111
00:05:26,333 --> 00:05:28,400
those cars will know everything

112
00:05:28,433 --> 00:05:31,233
about the road situation,
the road condition,

113
00:05:31,266 --> 00:05:33,066
well before a human would.

114
00:05:33,100 --> 00:05:37,833
So that if there is somebody
running around the corner

115
00:05:37,866 --> 00:05:39,633
about to jump in front
of the car,

116
00:05:39,666 --> 00:05:42,333
the car will know that,
and there will be no accidents.

117
00:05:42,366 --> 00:05:44,133
This is the dream.

118
00:05:44,166 --> 00:05:48,433
NARRATOR:
And that dream is
about much more than safety.

119
00:05:48,466 --> 00:05:51,766
Proponents say self-driving cars
could bring about

120
00:05:51,800 --> 00:05:53,933
the biggest changes
in transportation

121
00:05:53,966 --> 00:05:56,700
since horses gave way
to the automobile.

122
00:05:56,733 --> 00:05:59,833
Instead of having to own a car,

123
00:05:59,866 --> 00:06:02,966
people could simply summon one
from a circulating fleet

124
00:06:03,000 --> 00:06:04,733
of robotaxis,

125
00:06:04,766 --> 00:06:08,133
reducing the number of cars
on the road,

126
00:06:08,166 --> 00:06:11,000
cutting pollution,

127
00:06:11,033 --> 00:06:13,966
and eliminating the need
to have so many private cars

128
00:06:14,000 --> 00:06:16,733
sitting idle all day long.

129
00:06:16,766 --> 00:06:19,633
For people who can't drive,

130
00:06:19,666 --> 00:06:23,700
self-driving cars could also
provide greater mobility.

131
00:06:23,733 --> 00:06:26,666
SHAI SHALEV-SHWARTZ:
You can put your kid
on an autonomous car,

132
00:06:26,700 --> 00:06:30,033
and it will take him to school,
and problem solved.

133
00:06:30,066 --> 00:06:34,200
NARRATOR:
Some see a future
where cars talk to each other,

134
00:06:34,233 --> 00:06:35,233
reducing traffic jams.

135
00:06:35,266 --> 00:06:37,133
Others are less sure.

136
00:06:37,166 --> 00:06:38,433
GERDES:
It could also

137
00:06:38,466 --> 00:06:40,633
lead to the nightmare scenario,
as well,

138
00:06:40,666 --> 00:06:42,866
where inexpensive mobility

139
00:06:42,900 --> 00:06:45,466
leads to dramatic consumption
of mobility,

140
00:06:45,500 --> 00:06:48,933
congested freeways,
unsustainable use of energy,

141
00:06:48,966 --> 00:06:52,133
and an acceleration
of climate change

142
00:06:52,166 --> 00:06:54,433
and other issues
that we're facing now.

143
00:06:54,466 --> 00:06:56,200
SHASHUA:
This is a disruption.

144
00:06:56,233 --> 00:06:59,333
I would call this
"Mobility 2.0."

145
00:06:59,366 --> 00:07:01,566
If, if cars today is 1.0,

146
00:07:01,600 --> 00:07:06,033
and, uh, horse carriages
was 0.0,

147
00:07:06,066 --> 00:07:10,533
this is a new era of mobility.

148
00:07:10,566 --> 00:07:16,033
NARRATOR:
An era which in some places
seems very close at hand.

149
00:07:16,066 --> 00:07:20,433
Like here, on the streets
of San Francisco.

150
00:07:20,466 --> 00:07:21,900
All right,
I'm going to go ahead

151
00:07:21,933 --> 00:07:24,166
and slide to go.

152
00:07:24,200 --> 00:07:26,200
So we're off
on our autonomous way.

153
00:07:26,233 --> 00:07:28,233
NARRATOR:
Jesse Levinson demonstrates

154
00:07:28,266 --> 00:07:30,666
how his company's
self-driving car

155
00:07:30,700 --> 00:07:32,833
navigates the city streets.

156
00:07:32,866 --> 00:07:34,133
LEVINSON:
On the screen here,

157
00:07:34,166 --> 00:07:36,333
you can see what the vehicle
is planning to do.

158
00:07:36,366 --> 00:07:37,866
That's the green corridor.

159
00:07:37,900 --> 00:07:41,133
NARRATOR:
The display shows how
the route is constantly adjusted

160
00:07:41,166 --> 00:07:43,000
in response to data from cameras

161
00:07:43,033 --> 00:07:46,500
and scanners that use radar
and lasers.

162
00:07:46,533 --> 00:07:50,600
The system is designed to handle
anything that might happen.

163
00:07:50,633 --> 00:07:55,533
But on test drives, the company
always has a safety driver.

164
00:07:55,566 --> 00:07:56,900
Here's a really interesting
situation

165
00:07:56,933 --> 00:07:58,500
we're about to encounter,

166
00:07:58,533 --> 00:08:00,533
is a six-way
unprotected intersection.

167
00:08:00,566 --> 00:08:02,266
This intersection is
so complicated,

168
00:08:02,300 --> 00:08:04,200
I'm not sure
I know how to drive it.

169
00:08:04,233 --> 00:08:06,433
But what we're doing here is,
we're going to make a left turn.

170
00:08:06,466 --> 00:08:09,066
We're checking
all the oncoming traffic.

171
00:08:09,100 --> 00:08:11,100
NARRATOR:
The car scans its surroundings

172
00:08:11,133 --> 00:08:14,166
to determine
if its path is clear.

173
00:08:14,200 --> 00:08:16,633
LEVINSON:
We're also yielding
for all these pedestrians

174
00:08:16,666 --> 00:08:18,166
in the crosswalk.

175
00:08:18,200 --> 00:08:20,300
NARRATOR:
The computer won't allow the car
to proceed

176
00:08:20,333 --> 00:08:22,300
until it's safe.

177
00:08:22,333 --> 00:08:25,366
LEVINSON:
And literally tracking
hundreds of dynamic agents

178
00:08:25,400 --> 00:08:27,366
at the same time.

179
00:08:27,400 --> 00:08:29,400
Now, we've just made
our way back.

180
00:08:29,433 --> 00:08:30,566
That was a 100% autonomous drive

181
00:08:30,600 --> 00:08:32,833
with absolutely
no manual interventions.

182
00:08:32,866 --> 00:08:34,466
Pretty cool, huh?

183
00:08:34,500 --> 00:08:37,966
NARRATOR:
Test drives like this
are impressive.

184
00:08:38,000 --> 00:08:41,033
But some warn
it will be a long time

185
00:08:41,066 --> 00:08:44,300
before computers can
consistently drive more safely

186
00:08:44,333 --> 00:08:46,466
than humans do.

187
00:08:46,500 --> 00:08:48,966
Because driving,

188
00:08:49,000 --> 00:08:51,033
though it may seem easy,

189
00:08:51,066 --> 00:08:53,966
is actually
a very difficult task.

190
00:08:54,000 --> 00:08:56,233
RAJKUMAR:
Let me make the following
bold assertion.

191
00:08:56,266 --> 00:09:01,033
Driving is the most complex
activity

192
00:09:01,066 --> 00:09:04,300
that most adults on the planet
engage in

193
00:09:04,333 --> 00:09:05,866
on a regular basis.

194
00:09:05,900 --> 00:09:08,900
When we drive,
the vehicle is moving,

195
00:09:08,933 --> 00:09:11,333
the environment is changing
on a continuing basis,

196
00:09:11,366 --> 00:09:14,500
all these pieces of information
actually coming

197
00:09:14,533 --> 00:09:16,066
towards our senses,
goes to the brain,

198
00:09:16,100 --> 00:09:19,466
The brain basically applies
the rules of the road.

199
00:09:19,500 --> 00:09:21,900
And for the most part,
we drive safely.

200
00:09:21,933 --> 00:09:26,733
NARRATOR:
Each of the 35,000 annual
crash deaths in the U.S.

201
00:09:26,766 --> 00:09:28,266
is tragic.

202
00:09:28,300 --> 00:09:30,666
But they're statistically rare.

203
00:09:30,700 --> 00:09:32,666
On average, there's only one

204
00:09:32,700 --> 00:09:36,166
for every 100 million miles
of driving.

205
00:09:36,200 --> 00:09:41,300
That translates into
3.4 million hours of driving.

206
00:09:41,333 --> 00:09:46,900
3.4 million hours is 390 years

207
00:09:46,933 --> 00:09:50,533
of continuous, 24-hours-a-day,
seven-days-a-week driving

208
00:09:50,566 --> 00:09:52,133
in between fatal crashes.

209
00:09:52,166 --> 00:09:56,366
NARRATOR:
That's a very high bar
for technology to clear.

210
00:09:56,400 --> 00:09:59,133
Think about our modern
electronic devices

211
00:09:59,166 --> 00:10:01,633
that are powered by software
that we use every day.

212
00:10:01,666 --> 00:10:04,500
And try to imagine
that those devices could run

213
00:10:04,533 --> 00:10:08,433
without ever giving you
the spinning blue doughnut

214
00:10:08,466 --> 00:10:10,933
that said it's not ready to give
you the answer you wanted.

215
00:10:10,966 --> 00:10:14,866
Because if that computer
was driving your vehicle,

216
00:10:14,900 --> 00:10:16,233
you crashed.

217
00:10:16,266 --> 00:10:18,800
So getting to the point

218
00:10:18,833 --> 00:10:22,033
where we have
a software-intensive device

219
00:10:22,066 --> 00:10:26,100
that can operate without a fault

220
00:10:26,133 --> 00:10:28,266
is a huge, huge challenge.

221
00:10:28,300 --> 00:10:30,066
MICHAEL FLEMING:
This is a bold and ambitious
mission.

222
00:10:30,100 --> 00:10:32,033
A mission that really
isn't going

223
00:10:32,066 --> 00:10:33,433
to be accomplished overnight.

224
00:10:33,466 --> 00:10:37,833
NARRATOR:
Michael Fleming speaks
from hard experience.

225
00:10:37,866 --> 00:10:39,966
He's been working
on self-driving cars

226
00:10:40,000 --> 00:10:41,800
for more than a decade.

227
00:10:41,833 --> 00:10:44,300
And like many in the field
today,

228
00:10:44,333 --> 00:10:48,633
he got his start thanks to
a push from a surprising place:

229
00:10:48,666 --> 00:10:50,100
the Pentagon.

230
00:10:50,133 --> 00:10:54,333
In 2000, hoping to reduce
battlefield casualties,

231
00:10:54,366 --> 00:10:58,200
Congress orders the military
to develop combat vehicles

232
00:10:58,233 --> 00:11:00,533
that can drive themselves.

233
00:11:00,566 --> 00:11:02,766
Two years later,

234
00:11:02,800 --> 00:11:05,433
the Pentagon's research agency,
DARPA,

235
00:11:05,466 --> 00:11:09,666
announces what it calls
the Grand Challenge:

236
00:11:09,700 --> 00:11:12,266
a driverless car race,

237
00:11:12,300 --> 00:11:16,600
142 miles through
the California desert.

238
00:11:16,633 --> 00:11:20,800
Whoever finishes first
will win $1 million.

239
00:11:20,833 --> 00:11:24,466
March 13, 2004.

240
00:11:24,500 --> 00:11:27,966
13 vehicles-- rigged with
sensors to detect what's ahead

241
00:11:28,000 --> 00:11:30,266
and software to control speed
and steering--

242
00:11:30,300 --> 00:11:32,000
set out.

243
00:11:32,033 --> 00:11:34,266
FLEMING:
And we and a lot of other teams

244
00:11:34,300 --> 00:11:36,766
went out to compete in the
DARPA, you know, Grand Challenge

245
00:11:36,800 --> 00:11:40,133
with high hopes of bringing home
a million-dollar prize.

246
00:11:40,166 --> 00:11:43,833
NARRATOR:
Right away, the vehicles run
into trouble.

247
00:11:43,866 --> 00:11:46,933
The course is littered
with rocks, cliffs,

248
00:11:46,966 --> 00:11:51,166
cattle grates, and river beds--

249
00:11:51,200 --> 00:11:54,666
obstacles that the sensors
sometimes miss.

250
00:11:54,700 --> 00:11:56,733
We got 100 yards
out of the gate,

251
00:11:56,766 --> 00:11:58,633
and the vehicle
just stopped working.

252
00:11:58,666 --> 00:12:02,200
And we failed miserably
with everyone else.

253
00:12:02,233 --> 00:12:05,533
NARRATOR:
Of the full 142 miles,

254
00:12:05,566 --> 00:12:08,966
no vehicle goes farther
than eight.

255
00:12:09,000 --> 00:12:14,300
But a year later,
DARPA provides a second chance:

256
00:12:14,333 --> 00:12:16,233
a new challenge
that doubles the payoff

257
00:12:16,266 --> 00:12:18,566
to $2 million.

258
00:12:18,600 --> 00:12:23,466
Among the 2005 competitors,
a team from Stanford,

259
00:12:23,500 --> 00:12:27,200
confident that their car,
named Stanley,

260
00:12:27,233 --> 00:12:29,966
can make it all the way
to the finish line.

261
00:12:30,000 --> 00:12:31,700
Their ace in the hole:

262
00:12:31,733 --> 00:12:37,366
cutting-edge software that uses
A.I.-- artificial intelligence.

263
00:12:37,400 --> 00:12:38,900
SEBASTIAN THRUN:
We built a computer system

264
00:12:38,933 --> 00:12:40,433
using artificial intelligence

265
00:12:40,466 --> 00:12:43,833
that's able to actually find
the road very, very reliably.

266
00:12:43,866 --> 00:12:46,266
What you see here is
a map of the environment,

267
00:12:46,300 --> 00:12:47,866
that's being where
the Stanley drives.

268
00:12:47,900 --> 00:12:50,266
The red stuff is stuff that
it doesn't want to drive over,

269
00:12:50,300 --> 00:12:51,466
it's dangerous.

270
00:12:51,500 --> 00:12:53,366
The white stuff is the road
as found by Stanley,

271
00:12:53,400 --> 00:12:55,133
and the gray stuff
that you see here

272
00:12:55,166 --> 00:12:56,800
is stuff it just doesn't know
anything about,

273
00:12:56,833 --> 00:12:58,333
so it's not going to drive
there.

274
00:12:58,366 --> 00:12:59,900
(cheering and applauding)

275
00:12:59,933 --> 00:13:03,433
NARRATOR:
Stanford's strategy pays off.

276
00:13:03,466 --> 00:13:08,366
Stanley is the first to cross
the finish line.

277
00:13:08,400 --> 00:13:09,333
We have the green flag.

278
00:13:09,366 --> 00:13:11,233
NARRATOR:
Two years later...

279
00:13:11,266 --> 00:13:13,433
Launch the Bots!

280
00:13:13,466 --> 00:13:15,900
NARRATOR:
...the Pentagon stages
a final contest,

281
00:13:15,933 --> 00:13:17,866
adding a new complexity:

282
00:13:17,900 --> 00:13:19,766
traffic.

283
00:13:19,800 --> 00:13:23,266
DARPA calls it
the Urban Challenge.

284
00:13:23,300 --> 00:13:26,566
LEVINSON:
In the 2007 Urban Challenge,

285
00:13:26,600 --> 00:13:28,066
they let us drive
with other cars,

286
00:13:28,100 --> 00:13:31,133
both autonomous cars
as well as human-driven cars.

287
00:13:31,166 --> 00:13:32,866
And so that was really exciting,

288
00:13:32,900 --> 00:13:35,333
because all of a sudden, now you
have to track dynamic objects

289
00:13:35,366 --> 00:13:38,100
and predict what they're going
to do in the future,

290
00:13:38,133 --> 00:13:39,700
and that's
a much harder problem.

291
00:13:39,733 --> 00:13:41,100
NARRATOR:
To help solve it,

292
00:13:41,133 --> 00:13:43,833
some half a dozen teams bank
on a detection technology

293
00:13:43,866 --> 00:13:45,633
called lidar,

294
00:13:45,666 --> 00:13:49,600
which dramatically improves
a car's ability to see.

295
00:13:49,633 --> 00:13:53,833
They rig devices
that spin 360 degrees.

296
00:13:53,866 --> 00:13:56,666
Lidar works using laser beams,

297
00:13:56,700 --> 00:13:58,233
pulses of invisible light

298
00:13:58,266 --> 00:14:00,766
that bounce off everything
in their path.

299
00:14:00,800 --> 00:14:02,700
A sensor collects
the reflections,

300
00:14:02,733 --> 00:14:05,800
which provide a precise picture
of the environment.

301
00:14:05,833 --> 00:14:10,300
RYAN CHILTON:
And as those lasers are sweeping
around in a circle,

302
00:14:10,333 --> 00:14:12,366
you get what's called
a point cloud.

303
00:14:12,400 --> 00:14:14,566
So you'll get thousands,
even millions of points

304
00:14:14,600 --> 00:14:16,133
coming back to the sensor.

305
00:14:16,166 --> 00:14:17,366
And you can build up
a point cloud

306
00:14:17,400 --> 00:14:18,600
that looks similar to this,

307
00:14:18,633 --> 00:14:21,600
and it gives you
very accurate geometric detail.

308
00:14:21,633 --> 00:14:24,366
NARRATOR:
Using all that data
to pilot a car

309
00:14:24,400 --> 00:14:26,966
demands new kinds of software.

310
00:14:27,000 --> 00:14:29,500
DOLGOV:
I vividly remember
the first time

311
00:14:29,533 --> 00:14:33,000
when some of my software that
I had just written hours ago

312
00:14:33,033 --> 00:14:34,100
ran on the car.

313
00:14:34,133 --> 00:14:35,700
That was pretty incredible.

314
00:14:35,733 --> 00:14:37,200
There was nobody
behind the wheel,

315
00:14:37,233 --> 00:14:38,433
and it was just doing everything
on its own.

316
00:14:38,466 --> 00:14:40,400
NARRATOR:
But not quite perfectly.

317
00:14:40,433 --> 00:14:44,500
ANNOUNCER:
Okay, folks, we have got our
first autonomous traffic jam.

318
00:14:44,533 --> 00:14:47,100
ANNOUNCER 2:
Another historic event
right here.

319
00:14:47,133 --> 00:14:49,233
NARRATOR:
This time,

320
00:14:49,266 --> 00:14:53,133
six of 11 cars complete
the course successfully.

321
00:14:53,166 --> 00:14:55,000
It's a major turning point,

322
00:14:55,033 --> 00:14:58,333
but there's also
a growing appreciation

323
00:14:58,366 --> 00:14:59,933
for the immense challenge ahead.

324
00:14:59,966 --> 00:15:02,033
LEVINSON:
The reality was,
the Urban Challenge

325
00:15:02,066 --> 00:15:03,666
was a very small step

326
00:15:03,700 --> 00:15:04,933
compared to what had to be done

327
00:15:04,966 --> 00:15:06,833
to actually get
commercial vehicles on the road.

328
00:15:06,866 --> 00:15:09,433
It was only a six-hour race.

329
00:15:09,466 --> 00:15:11,633
So basically, if your car
could last for six hours

330
00:15:11,666 --> 00:15:12,766
without hitting something,

331
00:15:12,800 --> 00:15:14,700
you were, like,
"Yep, we did it," right?

332
00:15:14,733 --> 00:15:16,966
Now, getting into
a commercial service

333
00:15:17,000 --> 00:15:18,566
with thousands of vehicles

334
00:15:18,600 --> 00:15:20,833
and setting a safety bar
that's significantly higher

335
00:15:20,866 --> 00:15:22,233
than human-level performance,

336
00:15:22,266 --> 00:15:23,600
that's very difficult.

337
00:15:23,633 --> 00:15:26,500
NARRATOR:
Still, the potential
of this new technology

338
00:15:26,533 --> 00:15:29,333
proves irresistible
to engineers

339
00:15:29,366 --> 00:15:31,466
and to business.

340
00:15:31,500 --> 00:15:33,233
♪

341
00:15:33,266 --> 00:15:35,733
The DARPA challenges
may have seemed

342
00:15:35,766 --> 00:15:37,866
like just a geeky
science project,

343
00:15:37,900 --> 00:15:41,800
but they trigger a race
to build a whole new industry.

344
00:15:41,833 --> 00:15:43,766
One of the first
out of the gate?

345
00:15:43,800 --> 00:15:45,933
Tech giant Google.

346
00:15:45,966 --> 00:15:47,300
Should we do
a simple test first?

347
00:15:47,333 --> 00:15:49,266
NARRATOR:
In university robotics labs...

348
00:15:49,300 --> 00:15:50,600
Yes!

349
00:15:50,633 --> 00:15:51,833
NARRATOR:
...and the start-ups

350
00:15:51,866 --> 00:15:53,733
that spin out of them,

351
00:15:53,766 --> 00:15:56,166
a growing army of engineers
keeps improving

352
00:15:56,200 --> 00:15:58,633
both sensors and software.

353
00:15:58,666 --> 00:16:01,233
The big car manufacturers
take notice,

354
00:16:01,266 --> 00:16:03,933
and they too enter the fray.

355
00:16:03,966 --> 00:16:06,966
MAN:
Well, hello,
ladies and gentlemen,

356
00:16:07,000 --> 00:16:08,333
and welcome to C.E.S.

357
00:16:08,366 --> 00:16:10,733
NARRATOR:
They're betting
that driverless cars

358
00:16:10,766 --> 00:16:14,366
are about to become
a huge global business.

359
00:16:14,400 --> 00:16:16,900
RAJKUMAR:
The autonomous vehicle market

360
00:16:16,933 --> 00:16:19,966
is supposed to be an
enormous market in the future.

361
00:16:20,000 --> 00:16:23,900
Estimates say
that in the mid-2030s,

362
00:16:23,933 --> 00:16:28,066
the market could be worth
a few trillion U.S. dollars,

363
00:16:28,100 --> 00:16:30,766
with a T in there.

364
00:16:30,800 --> 00:16:34,400
NARRATOR:
That potential payday
brings Uber to Arizona,

365
00:16:34,433 --> 00:16:38,600
whose flat landscape
and sunny weather are ideal

366
00:16:38,633 --> 00:16:41,133
for testing self-driving cars.

367
00:16:41,166 --> 00:16:44,500
Uber sees eliminating
its paid drivers

368
00:16:44,533 --> 00:16:47,700
as a key
to future profitability.

369
00:16:47,733 --> 00:16:51,733
But the March 2018 crash
in Tempe

370
00:16:51,766 --> 00:16:55,733
casts a dark shadow on the
future of self-driving cars.

371
00:16:55,766 --> 00:17:00,500
Uber suspends all testing
on public roads.

372
00:17:00,533 --> 00:17:03,333
The National Transportation
Safety Board

373
00:17:03,366 --> 00:17:05,500
launches an investigation:

374
00:17:05,533 --> 00:17:07,599
why did neither
the computer system

375
00:17:07,633 --> 00:17:09,133
nor the safety driver

376
00:17:09,166 --> 00:17:11,133
stop the car?

377
00:17:11,166 --> 00:17:13,200
PHILIP KOOPMAN:
The Uber crash in Tempe

378
00:17:13,233 --> 00:17:16,066
wasn't really about the maturity
of the technology.

379
00:17:16,099 --> 00:17:18,400
We all knew the
self-driving car technology

380
00:17:18,433 --> 00:17:19,800
wasn't ready to deploy.

381
00:17:19,833 --> 00:17:21,233
That's why they were doing
testing.

382
00:17:21,266 --> 00:17:23,500
The significance
of the Uber crash

383
00:17:23,533 --> 00:17:26,700
was that the human safety driver
failed to prevent the crash.

384
00:17:26,733 --> 00:17:29,266
NARRATOR:
In the months after the crash,

385
00:17:29,300 --> 00:17:33,600
the NTSB and Tempe police
release new details

386
00:17:33,633 --> 00:17:37,600
about what happened that night.

387
00:17:37,633 --> 00:17:39,066
The interior camera reveals

388
00:17:39,100 --> 00:17:42,166
that Rafaela Vasquez
is not watching the road

389
00:17:42,200 --> 00:17:45,833
for nearly seven of the
22 minutes before the crash,

390
00:17:45,866 --> 00:17:49,433
including about five
of the last six seconds.

391
00:17:49,466 --> 00:17:52,633
(Tish Haynes Keyes vocalizing)

392
00:17:52,666 --> 00:17:55,733
KEYES:
♪ Chain, chain, chain

393
00:17:55,766 --> 00:17:57,066
NARRATOR:
The police discover

394
00:17:57,100 --> 00:17:59,533
that the whole time
the car is moving,

395
00:17:59,566 --> 00:18:03,100
she is streaming an episode
of a singing competition

396
00:18:03,133 --> 00:18:04,233
on her phone.

397
00:18:04,266 --> 00:18:06,100
♪ Chain of fools

398
00:18:06,133 --> 00:18:09,100
NARRATOR:
Vasquez denies watching
the video

399
00:18:09,133 --> 00:18:11,066
or even looking at her phone.

400
00:18:11,100 --> 00:18:14,233
She says she was just checking
the control panel.

401
00:18:14,266 --> 00:18:16,500
But she does not step
on the brake

402
00:18:16,533 --> 00:18:20,266
until after the car
strikes Herzberg.

403
00:18:20,300 --> 00:18:23,866
The NTSB findings
also point to flaws

404
00:18:23,900 --> 00:18:25,800
in the self-driving system.

405
00:18:25,833 --> 00:18:28,233
The sensors actually detect
Herzberg

406
00:18:28,266 --> 00:18:30,633
six seconds before impact,

407
00:18:30,666 --> 00:18:33,566
but the system doesn't alert
the safety driver.

408
00:18:33,600 --> 00:18:35,600
It's not designed to.

409
00:18:35,633 --> 00:18:39,033
So even though their sensors
had detected this target--

410
00:18:39,066 --> 00:18:42,033
there was a potentially
hazardous situation--

411
00:18:42,066 --> 00:18:43,400
they made the decision

412
00:18:43,433 --> 00:18:46,466
to not give any of that
information to the test driver.

413
00:18:46,500 --> 00:18:48,200
I can't imagine why.

414
00:18:48,233 --> 00:18:53,233
NARRATOR:
And the computer doesn't decide
emergency braking is needed

415
00:18:53,266 --> 00:18:57,666
until 1.3 seconds before impact.

416
00:18:57,700 --> 00:19:00,633
But even then,
the car doesn't brake,

417
00:19:00,666 --> 00:19:02,666
because Uber has disabled

418
00:19:02,700 --> 00:19:07,233
the Volvo's factory-installed
emergency braking system.

419
00:19:07,266 --> 00:19:09,800
Uber wants to avoid jerky rides

420
00:19:09,833 --> 00:19:14,366
caused by unnecessary braking
for harmless objects.

421
00:19:14,400 --> 00:19:16,933
So the net result of that is,

422
00:19:16,966 --> 00:19:20,033
they took
a safe production vehicle

423
00:19:20,066 --> 00:19:23,733
and turned it
into an unsafe prototype

424
00:19:23,766 --> 00:19:26,500
that caused the pedestrian
to be killed.

425
00:19:26,533 --> 00:19:31,466
NARRATOR:
Uber tells the NTSB that
it's the safety driver's job

426
00:19:31,500 --> 00:19:36,133
to correct any mistakes made
by the self-driving system.

427
00:19:36,166 --> 00:19:38,033
BRYAN REIMER:
Now, while many people look

428
00:19:38,066 --> 00:19:39,900
at the safety driver
in that vehicle

429
00:19:39,933 --> 00:19:41,866
as, as perhaps the villain,

430
00:19:41,900 --> 00:19:44,333
the real issue is the system

431
00:19:44,366 --> 00:19:47,233
that that safety driver
was put into.

432
00:19:47,266 --> 00:19:50,900
They were put into a system
ripe for failure,

433
00:19:50,933 --> 00:19:53,833
where the expectations
of failure

434
00:19:53,866 --> 00:19:57,233
were far lower than
the actual probabilities.

435
00:19:57,266 --> 00:20:01,066
Perhaps because the computer
scientists and the engineers

436
00:20:01,100 --> 00:20:03,966
building these systems

437
00:20:04,000 --> 00:20:07,266
don't all appreciate
the complexities

438
00:20:07,300 --> 00:20:10,833
of, that human behavior brings
to the system.

439
00:20:10,866 --> 00:20:15,000
NARRATOR:
Uber declined to participate
in this film.

440
00:20:15,033 --> 00:20:17,733
But even before Tempe,

441
00:20:17,766 --> 00:20:21,533
there were signs
that drivers and automation

442
00:20:21,566 --> 00:20:25,933
don't always work together
very well.

443
00:20:25,966 --> 00:20:29,166
Automotive engineers
classify automation

444
00:20:29,200 --> 00:20:31,800
into levels, from zero to five.

445
00:20:31,833 --> 00:20:35,533
Fully automated cars,
the ones driven by computers,

446
00:20:35,566 --> 00:20:38,233
are levels three, four,
and five.

447
00:20:38,266 --> 00:20:42,833
Uber was testing its car
as a level four prototype.

448
00:20:42,866 --> 00:20:47,166
At the lower levels,
humans drive.

449
00:20:47,200 --> 00:20:51,233
In level zero cars,
they handle everything.

450
00:20:51,266 --> 00:20:52,933
Level one and two cars

451
00:20:52,966 --> 00:20:54,933
assist drivers
by regulating speed

452
00:20:54,966 --> 00:20:56,733
or keeping the car in its lane,

453
00:20:56,766 --> 00:20:58,766
or even both.

454
00:20:58,800 --> 00:21:01,300
They're partially automated.

455
00:21:01,333 --> 00:21:04,633
Partial automation is the
foundation for full automation.

456
00:21:04,666 --> 00:21:07,766
And how we respond and adapt

457
00:21:07,800 --> 00:21:10,200
to automation
at the lower levels

458
00:21:10,233 --> 00:21:13,700
gives us our window
into the future.

459
00:21:13,733 --> 00:21:18,366
NARRATOR:
Millions of level one and two
cars are on the road today.

460
00:21:18,400 --> 00:21:23,166
Insurance data suggest that
they are less likely to crash

461
00:21:23,200 --> 00:21:26,333
because most of them
automatically brake the car

462
00:21:26,366 --> 00:21:28,466
to avoid collisions.

463
00:21:28,500 --> 00:21:30,466
And indeed,
it appears these systems

464
00:21:30,500 --> 00:21:32,733
are making people
better drivers, because...

465
00:21:32,766 --> 00:21:36,700
Think of it
as an extra set of eyes and ears

466
00:21:36,733 --> 00:21:38,466
beyond your own eyes and ears.

467
00:21:38,500 --> 00:21:40,900
So you get that additional
vigilance

468
00:21:40,933 --> 00:21:42,266
from the sensor systems

469
00:21:42,300 --> 00:21:45,266
that may detect things
that you missed.

470
00:21:45,300 --> 00:21:49,766
NARRATOR:
But the added confidence
provided by partial automation

471
00:21:49,800 --> 00:21:53,233
also poses an unforeseen risk.

472
00:21:53,266 --> 00:21:55,733
REIMER:
In the Volvo,

473
00:21:55,766 --> 00:21:57,866
she's clearly using
pilot assist a lot.

474
00:21:57,900 --> 00:22:00,000
Which is great, keeping
her hands on the wheel,

475
00:22:00,033 --> 00:22:02,166
paying attention to
what's going on in front of her.

476
00:22:02,200 --> 00:22:03,266
Mm-hmm.

477
00:22:03,300 --> 00:22:05,366
NARRATOR:
A research team at M.I.T.

478
00:22:05,400 --> 00:22:07,533
is gathering data on how people
use partial automation.

479
00:22:07,566 --> 00:22:11,033
Seem to use Autopilot a lot?
Yeah, he's been using it

480
00:22:11,066 --> 00:22:12,533
a little bit more
than everyone else, actually.

481
00:22:12,566 --> 00:22:14,433
REIMER:
The cars we are studying today

482
00:22:14,466 --> 00:22:17,200
can automatically steer,
accelerate, and brake

483
00:22:17,233 --> 00:22:19,366
in ways that were not available
in production vehicles

484
00:22:19,400 --> 00:22:21,433
20 years ago.

485
00:22:21,466 --> 00:22:24,033
So I think the ultimate question
is, you know,

486
00:22:24,066 --> 00:22:26,266
"How does attention change
over time

487
00:22:26,300 --> 00:22:29,500
as we use these cars
over months and years?"

488
00:22:29,533 --> 00:22:32,400
And that's the very heart
of the question

489
00:22:32,433 --> 00:22:33,933
we're hoping to understand more.

490
00:22:33,966 --> 00:22:37,733
NARRATOR:
Cathy Urquhart was given
a Volvo S90 to test-drive.

491
00:22:37,766 --> 00:22:40,200
It's equipped with
several automated features,

492
00:22:40,233 --> 00:22:45,466
including one for steering
that keeps the car in its lane.

493
00:22:45,500 --> 00:22:46,900
The first day I was nervous,
of course.

494
00:22:46,933 --> 00:22:49,200
It's new technology,
and I think it's the same

495
00:22:49,233 --> 00:22:50,600
when you're driving any new car.

496
00:22:50,633 --> 00:22:52,533
You have to get used to it.

497
00:22:52,566 --> 00:22:55,566
But I liked the lane monitor to
keep you in the, in the lane.

498
00:22:55,600 --> 00:22:57,966
Uh, it wasn't too obtrusive

499
00:22:58,000 --> 00:23:00,300
with the, uh, car
just kind of pulling you over,

500
00:23:00,333 --> 00:23:01,766
the wheel pulling you
to one side.

501
00:23:01,800 --> 00:23:04,133
It was good, I liked that.

502
00:23:04,166 --> 00:23:06,666
NARRATOR:
When the Volvo's steering
assistance is on,

503
00:23:06,700 --> 00:23:09,933
Cathy could briefly
take her hands off the wheel--

504
00:23:09,966 --> 00:23:11,166
but she doesn't.

505
00:23:11,200 --> 00:23:13,533
URQUHART:
I'm not somebody
that takes my hands

506
00:23:13,566 --> 00:23:15,300
off the steering wheel--
I like to have control.

507
00:23:15,333 --> 00:23:16,733
I like to hold on
to the steering wheel.

508
00:23:16,766 --> 00:23:19,733
NARRATOR:
But there are others
in the M.I.T. study,

509
00:23:19,766 --> 00:23:21,733
like Taylor Ogan,

510
00:23:21,766 --> 00:23:25,133
who are much more trusting
of automation.

511
00:23:25,166 --> 00:23:29,566
He owns a Tesla equipped
with software called Autopilot.

512
00:23:29,600 --> 00:23:32,800
OGAN:
Actually traded my first car in
to get Autopilot,

513
00:23:32,833 --> 00:23:34,733
so that the car
could drive itself.

514
00:23:34,766 --> 00:23:38,000
It helps you stay in your lane

515
00:23:38,033 --> 00:23:39,200
and follow the car,

516
00:23:39,233 --> 00:23:42,300
the distance of the car
in front of you.

517
00:23:42,333 --> 00:23:44,200
And you really don't have
to touch anything.

518
00:23:44,233 --> 00:23:45,500
It's awesome, I love it.

519
00:23:45,533 --> 00:23:47,833
I swear by it.

520
00:23:47,866 --> 00:23:49,333
All right, here we go.
All right.

521
00:23:49,366 --> 00:23:51,433
All right, oh!
(laughing)

522
00:23:51,466 --> 00:23:54,766
NARRATOR:
Tesla's founder and C.E.O.,
Elon Musk,

523
00:23:54,800 --> 00:23:58,200
has long touted Autopilot
as a step along the road

524
00:23:58,233 --> 00:24:00,700
to his goal of full autonomy.

525
00:24:00,733 --> 00:24:03,333
No hands, no feet,
nothing.

526
00:24:03,366 --> 00:24:06,766
It's a combination
of radar,

527
00:24:06,800 --> 00:24:09,100
camera with
image recognition,

528
00:24:09,133 --> 00:24:10,433
and ultrasonic sensors

529
00:24:10,466 --> 00:24:13,600
that's integrated with maps
and real-time traffic.

530
00:24:13,633 --> 00:24:17,866
NARRATOR:
But for now, the Tesla system
is still level two,

531
00:24:17,900 --> 00:24:21,400
which means it requires
just as much driver attention

532
00:24:21,433 --> 00:24:24,066
as a regular car.

533
00:24:24,100 --> 00:24:26,000
So the company has been
criticized

534
00:24:26,033 --> 00:24:28,066
for choosing the name Autopilot.

535
00:24:28,100 --> 00:24:29,866
PETER NORTON:
It's a nice, catchy-sounding
name,

536
00:24:29,900 --> 00:24:34,133
but "Autopilot" certainly
has a misleading connotation.

537
00:24:34,166 --> 00:24:37,100
It suggests
that the pilot of the car

538
00:24:37,133 --> 00:24:39,233
is in the program,
in the computer,

539
00:24:39,266 --> 00:24:42,500
and not in the driver
of the vehicle.

540
00:24:42,533 --> 00:24:46,266
NARRATOR:
But Musk claims that
the name is not misleading.

541
00:24:46,300 --> 00:24:48,333
MUSK:
Autopilot is what
they have in airplanes.

542
00:24:48,366 --> 00:24:50,933
It's where there's still an
expectation there'll be a pilot.

543
00:24:50,966 --> 00:24:52,866
So the, so if, if,

544
00:24:52,900 --> 00:24:54,400
the, the onus is on the pilot
to make sure

545
00:24:54,433 --> 00:24:57,033
that the autopilot is doing
the right thing.

546
00:24:57,066 --> 00:25:00,533
We're not asserting that the car
is capable of driving, um,

547
00:25:00,566 --> 00:25:03,500
in the absence
of driver oversight.

548
00:25:03,533 --> 00:25:06,533
NARRATOR:
Yet when he uses Autopilot,

549
00:25:06,566 --> 00:25:08,733
Taylor Ogan feels free
to look away from the road

550
00:25:08,766 --> 00:25:09,833
from time to time.

551
00:25:09,866 --> 00:25:11,400
OGAN:
I'm not a distracted driver,

552
00:25:11,433 --> 00:25:13,133
but I definitely do things

553
00:25:13,166 --> 00:25:15,400
that I think older people
who don't trust technology

554
00:25:15,433 --> 00:25:16,833
wouldn't do.

555
00:25:16,866 --> 00:25:19,333
I'm definitely on my phone
a lot more,

556
00:25:19,366 --> 00:25:21,366
texting, reading articles,

557
00:25:21,400 --> 00:25:23,666
when the car is driving itself.

558
00:25:23,700 --> 00:25:26,700
UFUOMA ELVIS-OBUKOWHO
I'm actually going to test
this Autopilot thing.

559
00:25:26,733 --> 00:25:28,433
Oh, I can eat food!

560
00:25:28,466 --> 00:25:30,866
(with mouth full):
This is the life, look at this.

561
00:25:30,900 --> 00:25:33,000
♪

562
00:25:33,033 --> 00:25:36,900
NARRATOR:
Some early Autopilot users push
things a lot further

563
00:25:36,933 --> 00:25:39,833
and show off their exploits
on YouTube.

564
00:25:39,866 --> 00:25:40,800
Oh, my!

565
00:25:40,833 --> 00:25:41,733
That's a sharp turn.

566
00:25:41,766 --> 00:25:43,300
(laughing):
Oh!

567
00:25:43,333 --> 00:25:44,300
I'm alive.

568
00:25:44,333 --> 00:25:45,300
Okay, good.

569
00:25:45,333 --> 00:25:46,666
(laughing)

570
00:25:46,700 --> 00:25:47,900
What is technology?

571
00:25:47,933 --> 00:25:49,100
What! What!

572
00:25:49,133 --> 00:25:50,900
NARRATOR:
Drivers like these

573
00:25:50,933 --> 00:25:53,933
may seem hopelessly reckless.

574
00:25:53,966 --> 00:25:56,100
But they're just
extreme examples

575
00:25:56,133 --> 00:25:57,733
of people placing too much trust

576
00:25:57,766 --> 00:26:01,500
in partially automated cars,

577
00:26:01,533 --> 00:26:06,366
a problem that in 2016
leads to tragedy.

578
00:26:06,400 --> 00:26:09,533
JOSHUA BROWN:
Ah, jeez, car's doing it
all itself.

579
00:26:09,566 --> 00:26:11,366
Don't know what I'm going to do
with my hands down here.

580
00:26:11,400 --> 00:26:13,666
(chuckles)

581
00:26:13,700 --> 00:26:19,033
NARRATOR:
39-year-old Joshua Brown is
a particularly enthusiastic

582
00:26:19,066 --> 00:26:21,733
early adopter
of Tesla's Autopilot.

583
00:26:21,766 --> 00:26:26,733
He loves to make YouTube videos
showing what Autopilot can do.

584
00:26:26,766 --> 00:26:30,400
BROWN:
So the camera up here is
what's always watching the road,

585
00:26:30,433 --> 00:26:31,533
and it's looking
at your lane markings

586
00:26:31,566 --> 00:26:33,500
and it is looking
at speed limit signs,

587
00:26:33,533 --> 00:26:34,766
and right now we're actually
about...

588
00:26:34,800 --> 00:26:36,333
NARRATOR:
Tesla says that Autopilot

589
00:26:36,366 --> 00:26:40,300
should only be used on highways
with entrance and exit ramps.

590
00:26:40,333 --> 00:26:45,266
Highways that don't have
cross traffic, like this one.

591
00:26:45,300 --> 00:26:48,933
But it's only a recommendation,

592
00:26:48,966 --> 00:26:52,233
and Tesla doesn't block drivers

593
00:26:52,266 --> 00:26:53,600
from using it
on other roads, too.

594
00:26:53,633 --> 00:26:54,866
BROWN:
There's a car in front of me,

595
00:26:54,900 --> 00:26:56,100
so I'm not going to have
to do anything.

596
00:26:56,133 --> 00:26:57,300
It starts to turn and...

597
00:26:57,333 --> 00:26:58,700
NARRATOR:
Brown stresses how important
it is

598
00:26:58,733 --> 00:27:02,166
for the driver to stay
constantly alert.

599
00:27:02,200 --> 00:27:03,900
BROWN:
So just like that,
I have my hands on the wheel,

600
00:27:03,933 --> 00:27:05,933
but I can still get
my foot down to a brake,

601
00:27:05,966 --> 00:27:08,800
and just keep it like this,
because you can react so quickly

602
00:27:08,833 --> 00:27:10,633
that if anything goes wrong,
you're going to want

603
00:27:10,666 --> 00:27:12,866
to be able to take control
very, very quickly

604
00:27:12,900 --> 00:27:14,400
while we're driving like this.

605
00:27:16,733 --> 00:27:21,666
NARRATOR:
On May 7, 2016, near
the town of Williston, Florida,

606
00:27:21,700 --> 00:27:25,800
Joshua Brown is cruising east
on Route 27A,

607
00:27:25,833 --> 00:27:31,533
a four-lane divided highway
that has numerous intersections.

608
00:27:31,566 --> 00:27:34,333
He is driving
at 74 miles per hour

609
00:27:34,366 --> 00:27:37,000
with Autopilot switched on.

610
00:27:37,033 --> 00:27:39,500
As Brown nears an intersection,

611
00:27:39,533 --> 00:27:43,866
a truck driver prepares to make
a left turn across his path.

612
00:27:43,900 --> 00:27:45,833
The truck is supposed to yield.

613
00:27:45,866 --> 00:27:47,400
But it doesn't.

614
00:27:47,433 --> 00:27:49,066
It begins its turn

615
00:27:49,100 --> 00:27:52,166
with Brown's Tesla
about 1,000 feet away.

616
00:27:52,200 --> 00:27:54,033
Brown now has about ten seconds

617
00:27:54,066 --> 00:27:56,566
to avoid a collision.

618
00:27:56,600 --> 00:28:00,266
Records show
that he is not using his phone.

619
00:28:00,300 --> 00:28:02,166
But he does nothing.

620
00:28:02,200 --> 00:28:04,200
And neither does Autopilot.

621
00:28:04,233 --> 00:28:05,500
(car crashing)

622
00:28:05,533 --> 00:28:07,633
The Tesla slams
into the trailer,

623
00:28:07,666 --> 00:28:10,233
killing Brown.

624
00:28:10,266 --> 00:28:15,033
In the immediate aftermath,
a critically important question:

625
00:28:15,066 --> 00:28:19,333
who-- or what-- is to blame?

626
00:28:19,366 --> 00:28:22,100
Autopilot is an obvious suspect.

627
00:28:22,133 --> 00:28:25,733
But the NTSB clears it

628
00:28:25,766 --> 00:28:30,333
because Autopilot isn't
designed to detect obstacles,

629
00:28:30,366 --> 00:28:31,333
like that truck,

630
00:28:31,366 --> 00:28:35,033
that are crossing
the car's path.

631
00:28:35,066 --> 00:28:36,133
It only looks for objects

632
00:28:36,166 --> 00:28:40,366
moving in the same direction
as the car.

633
00:28:40,400 --> 00:28:42,600
So, the Safety Board
splits the blame

634
00:28:42,633 --> 00:28:46,366
between the truck driver
for failing to yield

635
00:28:46,400 --> 00:28:50,566
and Brown
for not paying attention.

636
00:28:50,600 --> 00:28:53,233
Had he noticed the truck
and stepped on the brake,

637
00:28:53,266 --> 00:28:58,800
he could have easily stopped
with plenty of room to spare.

638
00:28:58,833 --> 00:29:02,566
And Brown is not the only person
killed while using Autopilot.

639
00:29:05,233 --> 00:29:07,300
In 2016 in China,

640
00:29:07,333 --> 00:29:11,500
a driver named Gao Yaning
crashes into a road sweeper.

641
00:29:11,533 --> 00:29:14,100
REPORTER:
A deadly Tesla crash
in California.

642
00:29:14,133 --> 00:29:16,566
NARRATOR:
In 2018, in California,

643
00:29:16,600 --> 00:29:18,866
Walter Huang dies when his
Tesla crashes into a barrier.

644
00:29:18,900 --> 00:29:21,166
REPORTER:
...set on Autopilot...

645
00:29:21,200 --> 00:29:24,700
REPORTER:
This deadly Tesla crash raising
new questions tonight.

646
00:29:24,733 --> 00:29:25,833
REPORTER:
The roof of the car
was ripped off

647
00:29:25,866 --> 00:29:27,066
as it passed
under the trailer...

648
00:29:27,100 --> 00:29:29,900
NARRATOR:
And in 2019, Jeremy Beren Banner

649
00:29:29,933 --> 00:29:32,066
is killed in Florida

650
00:29:32,100 --> 00:29:35,233
when his Tesla hits
a truck crossing its path.

651
00:29:35,266 --> 00:29:36,533
REPORTER:
The circumstances
of which are similar

652
00:29:36,566 --> 00:29:40,600
to a crash in May of 2016
near Gainesville.

653
00:29:42,466 --> 00:29:44,000
NARRATOR:
Despite these fatalities,

654
00:29:44,033 --> 00:29:48,300
Tesla says that it continually
improves Autopilot.

655
00:29:48,333 --> 00:29:51,566
And it asserts that drivers
who use Autopilot

656
00:29:51,600 --> 00:29:56,200
have a lower crash rate
than U.S. drivers as a whole.

657
00:29:58,166 --> 00:30:00,833
Tesla declined to participate
in this film.

658
00:30:00,866 --> 00:30:04,466
Some scientists say
that Autopilot

659
00:30:04,500 --> 00:30:05,500
and other systems like it

660
00:30:05,533 --> 00:30:10,100
remain inherently risky.

661
00:30:10,133 --> 00:30:12,566
When cars are fully manual and
you must do everything yourself,

662
00:30:12,600 --> 00:30:14,800
you bring
all your cognitive resources

663
00:30:14,833 --> 00:30:16,933
to bear to do that task,

664
00:30:16,966 --> 00:30:19,466
but if the automation
is doing a good enough job,

665
00:30:19,500 --> 00:30:21,833
people will check out
in their heads very quickly.

666
00:30:21,866 --> 00:30:26,366
Your brain is wired
to stop paying attention

667
00:30:26,400 --> 00:30:30,400
when the automation
starts doing well enough.

668
00:30:30,433 --> 00:30:34,300
It's very easy to be lulled
into a false sense of security

669
00:30:34,333 --> 00:30:36,900
if you're supervising
an automated vehicle

670
00:30:36,933 --> 00:30:40,333
that you have never seen
have any sort of issue.

671
00:30:40,366 --> 00:30:41,833
(siren wailing)

672
00:30:41,866 --> 00:30:44,533
NARRATOR:
That false sense of security
makes it difficult

673
00:30:44,566 --> 00:30:47,400
for people to react quickly
in emergencies.

674
00:30:47,433 --> 00:30:51,433
So, many engineers are skeptical
of what's called

675
00:30:51,466 --> 00:30:53,433
level three autonomy,

676
00:30:53,466 --> 00:30:55,500
where a car can fully
drive itself,

677
00:30:55,533 --> 00:30:56,866
except in an emergency,

678
00:30:56,900 --> 00:31:00,000
when the system alerts
the driver to take over.

679
00:31:00,033 --> 00:31:02,133
LEVINSON:
If the vehicle
all of a sudden can say,

680
00:31:02,166 --> 00:31:03,400
"Oh, wait, I don't actually
know what I'm doing,

681
00:31:03,433 --> 00:31:05,066
you better take over,"

682
00:31:05,100 --> 00:31:06,500
I'm not at all convinced

683
00:31:06,533 --> 00:31:08,400
that that can be done
in a way that's actually safe.

684
00:31:08,433 --> 00:31:10,066
That's one of the reasons

685
00:31:10,100 --> 00:31:11,433
why we're making the leap
all the way

686
00:31:11,466 --> 00:31:13,900
to level four and five driving,
which is to say

687
00:31:13,933 --> 00:31:15,433
that as a passenger
in our vehicle,

688
00:31:15,466 --> 00:31:17,633
you have no legal
or other responsibility

689
00:31:17,666 --> 00:31:20,866
for driving
or keeping the vehicle safe.

690
00:31:20,900 --> 00:31:23,366
NARRATOR:
Level four and five cars,

691
00:31:23,400 --> 00:31:25,133
which are fully autonomous,

692
00:31:25,166 --> 00:31:27,933
are designed to safely handle
any driving situation,

693
00:31:27,966 --> 00:31:30,300
even emergencies,

694
00:31:30,333 --> 00:31:33,300
with no human intervention
at all.

695
00:31:33,333 --> 00:31:36,866
The only difference:
level fours could operate only

696
00:31:36,900 --> 00:31:38,133
on some roads at certain times;

697
00:31:38,166 --> 00:31:43,533
level fives, anytime, anywhere.

698
00:31:43,566 --> 00:31:44,566
SHASHUA:
You can go to sleep

699
00:31:44,600 --> 00:31:46,633
or sit in the back seat,

700
00:31:46,666 --> 00:31:48,266
because there is
no take-over request.

701
00:31:48,300 --> 00:31:52,533
The car can manage
all the situations.

702
00:31:52,566 --> 00:31:55,166
And in case something
goes wrong,

703
00:31:55,200 --> 00:31:56,966
it still has enough redundancies

704
00:31:57,000 --> 00:31:59,500
to safely stop.

705
00:31:59,533 --> 00:32:01,800
So it doesn't need
the driver to engage.

706
00:32:01,833 --> 00:32:05,400
NARRATOR:
But achieving that goal,
of fully driverless cars,

707
00:32:05,433 --> 00:32:09,266
will require
the people developing them

708
00:32:09,300 --> 00:32:12,033
to overcome
all kinds of obstacles.

709
00:32:12,066 --> 00:32:13,500
We've never done this before.

710
00:32:13,533 --> 00:32:15,033
We've done some things like it.

711
00:32:15,066 --> 00:32:17,933
We've increased safety
in aviation tremendously.

712
00:32:17,966 --> 00:32:19,566
We've automated
different kinds

713
00:32:19,600 --> 00:32:22,400
of transportation modes
and activities beautifully.

714
00:32:22,433 --> 00:32:24,033
But we've never done this.

715
00:32:24,066 --> 00:32:27,600
♪

716
00:32:27,633 --> 00:32:28,800
(whirring)

717
00:32:29,900 --> 00:32:31,733
GERDES:
Trying to develop
an automated vehicle

718
00:32:31,766 --> 00:32:33,900
that can do everything
that a human driver can do

719
00:32:33,933 --> 00:32:36,466
is a huge problem.

720
00:32:36,500 --> 00:32:39,166
And it requires an awful lot
of interlocking pieces.

721
00:32:39,200 --> 00:32:42,233
NARRATOR:
To even match human drivers,

722
00:32:42,266 --> 00:32:46,933
a self-driving car needs
to learn to handle not just one,

723
00:32:46,966 --> 00:32:49,900
but three distinct tasks
nearly flawlessly.

724
00:32:49,933 --> 00:32:53,866
The first: seeing everything
that's around the car.

725
00:32:53,900 --> 00:32:58,633
Second, understanding
what it's seeing.

726
00:32:58,666 --> 00:33:01,966
And third, planning
the car's path

727
00:33:02,000 --> 00:33:04,866
and controlling
its acceleration, braking,

728
00:33:04,900 --> 00:33:07,333
and steering along the way.

729
00:33:07,366 --> 00:33:10,566
First: seeing.

730
00:33:10,600 --> 00:33:14,866
Most self-driving cars rely
on cameras, radar, and lidar.

731
00:33:14,900 --> 00:33:17,200
Each has weaknesses.

732
00:33:17,233 --> 00:33:19,966
Cameras work poorly at night.

733
00:33:20,000 --> 00:33:22,366
Radar doesn't
distinguish very well

734
00:33:22,400 --> 00:33:24,300
between different types
of objects.

735
00:33:24,333 --> 00:33:30,133
And lidar can fail completely
in rain, snow, or fog.

736
00:33:30,166 --> 00:33:34,933
But even when sensors
are working nearly perfectly,

737
00:33:34,966 --> 00:33:37,066
there's still
the second big challenge:

738
00:33:37,100 --> 00:33:40,833
understanding
what they're seeing.

739
00:33:42,833 --> 00:33:45,733
For us, taking in
sensory information

740
00:33:45,766 --> 00:33:48,733
and making best guesses
about the real world

741
00:33:48,766 --> 00:33:49,933
are second nature.

742
00:33:49,966 --> 00:33:51,866
For a computer,

743
00:33:51,900 --> 00:33:55,033
making meaning
out of visual data

744
00:33:55,066 --> 00:33:56,900
is fiendishly difficult.

745
00:33:56,933 --> 00:34:01,100
In each case,
it's the task called perception.

746
00:34:01,133 --> 00:34:05,000
SHASHUA:
The human brain is an expert
in perception.

747
00:34:05,033 --> 00:34:07,333
But for machines,
it's not natural.

748
00:34:07,366 --> 00:34:09,033
So this is one big challenge,
right?

749
00:34:09,066 --> 00:34:10,933
If you want
to drive autonomously,

750
00:34:10,966 --> 00:34:14,033
you need to perceive the world
just like humans do.

751
00:34:14,066 --> 00:34:15,533
NARRATOR:
Until recently,

752
00:34:15,566 --> 00:34:18,000
no computer could come close.

753
00:34:18,033 --> 00:34:19,400
But since 2010,

754
00:34:19,433 --> 00:34:24,633
a big leap forward in a branch
of A.I. called machine learning

755
00:34:24,666 --> 00:34:26,666
has come a long way
to closing the gap.

756
00:34:26,699 --> 00:34:28,199
RUS:
Machine learning,

757
00:34:28,233 --> 00:34:32,866
it's about giving machines
the ability to look at data,

758
00:34:32,900 --> 00:34:34,866
identify patterns,
and make predictions.

759
00:34:34,900 --> 00:34:38,199
NARRATOR:
Some common examples
of machine learning:

760
00:34:38,233 --> 00:34:40,699
an app that can recognize
human speech,

761
00:34:40,733 --> 00:34:45,300
or an airport security system
that can recognize faces.

762
00:34:45,333 --> 00:34:48,100
But before a computer
can do these things,

763
00:34:48,133 --> 00:34:51,566
it has to be trained--
shown countless examples

764
00:34:51,600 --> 00:34:54,433
of the things
we want it to recognize.

765
00:34:54,466 --> 00:34:55,866
At the core of machine learning

766
00:34:55,900 --> 00:34:57,900
is the idea
of using training data

767
00:34:57,933 --> 00:35:00,900
so that one can train the system
to do something.

768
00:35:00,933 --> 00:35:03,366
Usually, the training stage

769
00:35:03,400 --> 00:35:05,366
consists of
millions of examples.

770
00:35:05,400 --> 00:35:10,100
NARRATOR:
In this case,
a set of images of cats.

771
00:35:10,133 --> 00:35:13,100
The computer will find
the similarities among them

772
00:35:13,133 --> 00:35:14,300
that will help it generalize

773
00:35:14,333 --> 00:35:18,000
and be able to recognize
any cat, like we do.

774
00:35:18,033 --> 00:35:22,233
RUS:
It also requires
a great breadth of examples,

775
00:35:22,266 --> 00:35:24,700
because the performance
of the network

776
00:35:24,733 --> 00:35:28,733
will be only as good
as the data used to train it.

777
00:35:30,133 --> 00:35:31,433
NARRATOR:
In Jerusalem,

778
00:35:31,466 --> 00:35:35,033
Mobileye, a specialist
in autonomous technology,

779
00:35:35,066 --> 00:35:39,833
trains its perception software
with a huge volume of data.

780
00:35:39,866 --> 00:35:44,100
A person has to carefully
label each image

781
00:35:44,133 --> 00:35:48,266
so that the computer can learn
what different things look like.

782
00:35:48,300 --> 00:35:51,600
She is making sure that all the
vehicle are correctly annotated.

783
00:35:51,633 --> 00:35:54,033
We are looking
for cars, pedestrian,

784
00:35:54,066 --> 00:35:56,300
traffic sign, traffic light.

785
00:35:56,333 --> 00:36:00,266
NARRATOR:
Once the software masters
the set of training images,

786
00:36:00,300 --> 00:36:03,400
it's ready to tackle the data
from a real drive,

787
00:36:03,433 --> 00:36:08,200
a stream of pixels coming in
from the car's cameras.

788
00:36:08,233 --> 00:36:10,566
KOOPMAN:
And it'll take each image
from the video,

789
00:36:10,600 --> 00:36:12,833
say, one frame out of the video,
and say,

790
00:36:12,866 --> 00:36:16,466
"Okay, I have a bunch of pixels,
a bunch of colored dots,"

791
00:36:16,500 --> 00:36:17,766
and it tries to find out
what's in there.

792
00:36:17,800 --> 00:36:19,900
And at a high level,
what it's doing is,

793
00:36:19,933 --> 00:36:21,533
it's going around,
sniffing through the image,

794
00:36:21,566 --> 00:36:22,533
looking for things like,

795
00:36:22,566 --> 00:36:24,133
"Oh, there are
some vertical edges,

796
00:36:24,166 --> 00:36:26,466
"and there are
some horizontal edges,

797
00:36:26,500 --> 00:36:27,933
and there's something round."

798
00:36:27,966 --> 00:36:29,933
And so it pulls out
a bunch of features.

799
00:36:29,966 --> 00:36:32,033
And so it's pulling
these video features out

800
00:36:32,066 --> 00:36:33,800
and associating them

801
00:36:33,833 --> 00:36:35,533
with whether or not
it's a person or a car.

802
00:36:35,566 --> 00:36:40,533
NARRATOR:
Today's software can interpret
millions of pixels every second.

803
00:36:40,566 --> 00:36:42,600
And thanks to the recent
breakthroughs

804
00:36:42,633 --> 00:36:44,933
in machine learning,

805
00:36:44,966 --> 00:36:46,666
its accuracy has shot up
substantially,

806
00:36:46,700 --> 00:36:48,966
to as high as 98%.

807
00:36:49,000 --> 00:36:52,466
But is that good enough?

808
00:36:52,500 --> 00:36:55,866
KOOPMAN:
Whenever I hear a good
high number like 98% accuracy

809
00:36:55,900 --> 00:36:58,833
in the context of
self-driving cars,

810
00:36:58,866 --> 00:37:01,366
my reaction is,
"That's not near good enough."

811
00:37:01,400 --> 00:37:03,933
One of the problems
with a really high accuracy is,

812
00:37:03,966 --> 00:37:06,766
that's only about
how you do on the training data.

813
00:37:06,800 --> 00:37:09,166
If the real world
is even slightly different

814
00:37:09,200 --> 00:37:11,533
than the training data--
which it always is--

815
00:37:11,566 --> 00:37:14,766
that accuracy might not
really turn out to be the case.

816
00:37:14,800 --> 00:37:16,866
NARRATOR:
In Pittsburgh,

817
00:37:16,900 --> 00:37:20,000
Phil Koopman's company
helps clients

818
00:37:20,033 --> 00:37:22,866
by pushing perception software
to its limits,

819
00:37:22,900 --> 00:37:26,666
hoping to reveal
the unexpected ways it can fail.

820
00:37:26,700 --> 00:37:27,733
He drops out.

821
00:37:27,766 --> 00:37:29,333
It picks him up
for just a little bit,

822
00:37:29,366 --> 00:37:30,466
and then he goes away.

823
00:37:30,500 --> 00:37:32,900
KOOPMAN:
The driverless cars have gotten

824
00:37:32,933 --> 00:37:35,366
really good
at ordinary situations.

825
00:37:35,400 --> 00:37:37,533
Going down the highway
on a sunny day

826
00:37:37,566 --> 00:37:38,700
should be no problem.

827
00:37:38,733 --> 00:37:41,333
Even navigating city streets,

828
00:37:41,366 --> 00:37:42,366
if nothing crazy's going on,

829
00:37:42,400 --> 00:37:44,566
should be okay.

830
00:37:44,600 --> 00:37:46,166
They have a lot of trouble

831
00:37:46,200 --> 00:37:48,033
with the things
they haven't seen before.

832
00:37:48,066 --> 00:37:49,166
So we call them edge cases.

833
00:37:49,200 --> 00:37:51,866
Just something
you've never seen before.

834
00:37:51,900 --> 00:37:52,966
JEN GALLINGANE:
And you'll notice

835
00:37:53,000 --> 00:37:54,666
it was raining a little bit
that night,

836
00:37:54,700 --> 00:37:56,266
so we have a lot of adults
carrying umbrellas.

837
00:37:56,300 --> 00:37:58,933
KOOPMAN:
If they don't have a lot of
umbrellas in the training set,

838
00:37:58,966 --> 00:38:00,400
maybe it's not going to see
the people with the umbrellas.

839
00:38:00,433 --> 00:38:01,500
It's not going to see
the people.

840
00:38:01,533 --> 00:38:02,566
KOOPMAN:
So when it sees something

841
00:38:02,600 --> 00:38:04,066
that it's never seen before--

842
00:38:04,100 --> 00:38:06,666
it's never seen an example--
it doesn't know what to do.

843
00:38:06,700 --> 00:38:10,666
NARRATOR:
Today, Koopman and his colleague
Jen Gallingane

844
00:38:10,700 --> 00:38:12,766
are driving around Pittsburgh
in his car,

845
00:38:12,800 --> 00:38:15,866
looking for edge cases.

846
00:38:15,900 --> 00:38:17,766
KOOPMAN:
We have the crossing guard
directing traffic

847
00:38:17,800 --> 00:38:19,766
in addition to a traffic light.

848
00:38:19,800 --> 00:38:21,966
NARRATOR:
They use what they learn

849
00:38:22,000 --> 00:38:23,666
to improve the training
of the software.

850
00:38:23,700 --> 00:38:27,500
KOOPMAN:
The red boxes you see
are called bounding boxes,

851
00:38:27,533 --> 00:38:29,300
and what those are,
they're just a rectangle

852
00:38:29,333 --> 00:38:31,033
around where the person is
in the image.

853
00:38:31,066 --> 00:38:32,866
And so the idea is,

854
00:38:32,900 --> 00:38:35,333
this is where the computer
thinks a pedestrian is.

855
00:38:35,366 --> 00:38:37,800
When you see
those bounding boxes disappear,

856
00:38:37,833 --> 00:38:39,166
it means that the computer
has lost track

857
00:38:39,200 --> 00:38:40,933
of where the person is.

858
00:38:40,966 --> 00:38:42,700
In other words,
if there's no red box,

859
00:38:42,733 --> 00:38:44,200
it doesn't see the person.

860
00:38:44,233 --> 00:38:46,033
So I don't think
it saw him hardly at all,

861
00:38:46,066 --> 00:38:47,400
until we were almost
on top of him.

862
00:38:47,433 --> 00:38:49,166
NARRATOR:
And there are many
common variations

863
00:38:49,200 --> 00:38:52,566
in the ways
that people and things look

864
00:38:52,600 --> 00:38:56,666
that can prevent a computer
from making the correct I.D.

865
00:38:56,700 --> 00:38:59,666
A delivery man
wearing a turban

866
00:38:59,700 --> 00:39:02,433
is holding a food tray
next to his head--

867
00:39:02,466 --> 00:39:05,266
forming shapes that differ
from the typical training images

868
00:39:05,300 --> 00:39:06,866
of a person's head.

869
00:39:06,900 --> 00:39:08,466
Got the food tray
right here.

870
00:39:08,500 --> 00:39:14,166
NARRATOR:
Those unusual shapes make
the delivery man an edge case.

871
00:39:14,200 --> 00:39:15,566
So when the red box is there,
it sees him,

872
00:39:15,600 --> 00:39:17,500
and there he is
looking right at me.

873
00:39:17,533 --> 00:39:20,233
He's right in front of my car,
it doesn't see him.

874
00:39:20,266 --> 00:39:21,366
KOOPMAN:
You can look at the picture,

875
00:39:21,400 --> 00:39:22,800
and you say,
"There's a person there."

876
00:39:22,833 --> 00:39:23,800
And the perception system says,

877
00:39:23,833 --> 00:39:25,100
"Nope, there's no person there."

878
00:39:25,133 --> 00:39:26,066
And that's a problem.

879
00:39:26,100 --> 00:39:28,100
NARRATOR:
But many engineers say

880
00:39:28,133 --> 00:39:31,066
that a botched identification
can be overcome--

881
00:39:31,100 --> 00:39:33,566
as long as
some of the car's sensors

882
00:39:33,600 --> 00:39:36,366
detect that something is there.

883
00:39:36,400 --> 00:39:38,566
So we use sensors
like radar and lidar

884
00:39:38,600 --> 00:39:40,433
to directly measure
where everything is around us.

885
00:39:40,466 --> 00:39:41,900
And that's really important,

886
00:39:41,933 --> 00:39:43,266
because even if
a machine learning system

887
00:39:43,300 --> 00:39:47,200
can't classify exactly what type
of object something is,

888
00:39:47,233 --> 00:39:49,066
we still know
there's something there,

889
00:39:49,100 --> 00:39:50,400
we know how fast it's moving,

890
00:39:50,433 --> 00:39:52,200
and so we can make sure
that we don't hit it.

891
00:39:52,233 --> 00:39:56,633
NARRATOR:
Even if the car can see and
make sense of its surroundings

892
00:39:56,666 --> 00:39:58,433
with unmatched perfection,

893
00:39:58,466 --> 00:40:01,400
it could still be
a lethal hazard

894
00:40:01,433 --> 00:40:04,833
if it fails to handle
its third crucial task:

895
00:40:04,866 --> 00:40:06,533
planning.

896
00:40:06,566 --> 00:40:08,666
Yet another daunting challenge.

897
00:40:08,700 --> 00:40:11,766
Because the planning software
has to anticipate

898
00:40:11,800 --> 00:40:13,433
what's likely to happen,

899
00:40:13,466 --> 00:40:16,100
then plot the car's pathway
and speed,

900
00:40:16,133 --> 00:40:18,600
ready to change both
in a split second

901
00:40:18,633 --> 00:40:21,666
if the sensors spot trouble.

902
00:40:21,700 --> 00:40:24,333
Just like the car's
perception software,

903
00:40:24,366 --> 00:40:27,466
its planning software
also needs to be trained.

904
00:40:27,500 --> 00:40:30,633
To do that
without putting lives at risk,

905
00:40:30,666 --> 00:40:32,600
most companies do
a lot of their training

906
00:40:32,633 --> 00:40:36,400
in environments
they can fully control.

907
00:40:36,433 --> 00:40:39,133
One of them
is computer simulation.

908
00:40:39,166 --> 00:40:42,433
We are driving in San Francisco,
so we have to create

909
00:40:42,466 --> 00:40:46,866
a world in simulation that
is just as complex and varied

910
00:40:46,900 --> 00:40:48,133
as San Francisco.

911
00:40:48,166 --> 00:40:50,566
And this is not a small feat.

912
00:40:50,600 --> 00:40:57,400
When we run tests in simulation,
we take a model of the real car,

913
00:40:57,433 --> 00:41:01,466
along with all the sensors
that are on the real car,

914
00:41:01,500 --> 00:41:04,266
including the A.I.
that runs on the real car.

915
00:41:04,300 --> 00:41:06,866
And this is what we place
in our simulation.

916
00:41:06,900 --> 00:41:09,366
NARRATOR:
Here, the A.I. software
can practice new skills

917
00:41:09,400 --> 00:41:12,066
without putting anyone
in danger.

918
00:41:12,100 --> 00:41:14,966
TARALOVA:
Suppose that we drive
in the real world,

919
00:41:15,000 --> 00:41:18,066
and there is a
double-parked car situation

920
00:41:18,100 --> 00:41:19,266
we don't know
how to deal with, right?

921
00:41:19,300 --> 00:41:21,633
So what we do is, we ensure

922
00:41:21,666 --> 00:41:23,533
that the vehicle can deal
with these situations

923
00:41:23,566 --> 00:41:24,833
in simulation first,

924
00:41:24,866 --> 00:41:27,200
so that once we actually
see that situation in real life,

925
00:41:27,233 --> 00:41:29,333
we already know that
we'll be able to deal with it.

926
00:41:29,366 --> 00:41:32,766
NARRATOR:
Another kind
of safe training environment

927
00:41:32,800 --> 00:41:34,066
is a private test facility,

928
00:41:34,100 --> 00:41:36,933
like this one
in northern California,

929
00:41:36,966 --> 00:41:38,200
known as Castle.

930
00:41:38,233 --> 00:41:39,633
It's run by Waymo,

931
00:41:39,666 --> 00:41:44,400
the company that Google spun off
to develop self-driving cars.

932
00:41:44,433 --> 00:41:48,633
And so, because the site has so
many different types of roads,

933
00:41:48,666 --> 00:41:52,100
from residential to expressways

934
00:41:52,133 --> 00:41:55,600
to arterial roads
going between the two,

935
00:41:55,633 --> 00:41:57,066
cul-de-sacs
and things like that,

936
00:41:57,100 --> 00:42:00,566
we're able to stage
basically an infinite number

937
00:42:00,600 --> 00:42:02,033
of scenarios
that you would encounter

938
00:42:02,066 --> 00:42:05,166
on those types of roads
in the real world.

939
00:42:05,200 --> 00:42:07,566
Great, let's go for run.

940
00:42:07,600 --> 00:42:10,933
(on radio):
Three, two, one.

941
00:42:12,566 --> 00:42:14,533
(horn honks)

942
00:42:14,566 --> 00:42:16,866
NARRATOR:
Today's test:

943
00:42:16,900 --> 00:42:19,500
on a street
where the view is blocked,

944
00:42:19,533 --> 00:42:22,400
a car suddenly backs out
into the Waymo car's path.

945
00:42:23,800 --> 00:42:24,800
TONY KUM (on radio):
Great job, everyone.

946
00:42:24,833 --> 00:42:27,133
We can go a little
spicier.

947
00:42:27,166 --> 00:42:28,500
VILLEGAS:
We can change the speed

948
00:42:28,533 --> 00:42:31,400
at which the auxiliary vehicle
exits the driveway.

949
00:42:31,433 --> 00:42:32,933
Does it rip out of its driveway

950
00:42:32,966 --> 00:42:35,100
like a bat out of hell, really
coming out of the driveway

951
00:42:35,133 --> 00:42:36,233
ahead of the Waymo vehicle?

952
00:42:36,266 --> 00:42:37,466
Or is it slowly kind of

953
00:42:37,500 --> 00:42:39,933
meandering down the driveway

954
00:42:39,966 --> 00:42:42,700
and taking its time?

955
00:42:42,733 --> 00:42:43,666
Great job, guys.

956
00:42:43,700 --> 00:42:45,100
Let's restage and stand by.

957
00:42:45,133 --> 00:42:46,666
NARRATOR:
Next,

958
00:42:46,700 --> 00:42:50,533
the Waymo car has to handle
what's called a pinch point:

959
00:42:50,566 --> 00:42:53,600
a two-way street made narrow
by parked cars.

960
00:42:53,633 --> 00:42:55,766
Waymo rolling.

961
00:42:55,800 --> 00:42:58,900
VILLEGAS:
You encounter this really often
on public roads:

962
00:42:58,933 --> 00:43:01,533
Two oncoming vehicles
have to negotiate

963
00:43:01,566 --> 00:43:02,866
who will assume
the right of way

964
00:43:02,900 --> 00:43:04,000
and who will have to yield.

965
00:43:04,033 --> 00:43:06,700
NARRATOR:
The other car arrives

966
00:43:06,733 --> 00:43:08,533
at the pinch point first,

967
00:43:08,566 --> 00:43:11,866
so the software should tell
the autonomous car to yield.

968
00:43:11,900 --> 00:43:14,933
GUNJAN YAGNIK:
Yielding, yielding right here.

969
00:43:14,966 --> 00:43:17,333
Like butter, going around.

970
00:43:17,366 --> 00:43:19,100
Butter.

971
00:43:19,133 --> 00:43:20,166
NARRATOR:
These are just two

972
00:43:20,200 --> 00:43:22,433
of the thousands of scenarios

973
00:43:22,466 --> 00:43:24,600
the company uses
to train the software,

974
00:43:24,633 --> 00:43:28,366
and they introduce new ones
all the time.

975
00:43:28,400 --> 00:43:33,300
But ultimately,
the only way to know for sure

976
00:43:33,333 --> 00:43:35,666
how a car will perform
on public roads

977
00:43:35,700 --> 00:43:37,833
is to test it
in the real world--

978
00:43:37,866 --> 00:43:40,666
where the stakes
for wrong decisions

979
00:43:40,700 --> 00:43:43,433
are much higher.

980
00:43:43,466 --> 00:43:46,033
We don't just,
"Let's go for a test drive

981
00:43:46,066 --> 00:43:47,933
just for the,
for the fun of it."

982
00:43:47,966 --> 00:43:51,700
Because a test drive by itself
may be dangerous, okay?

983
00:43:51,733 --> 00:43:56,566
But there are things that are
very difficult to check

984
00:43:56,600 --> 00:43:58,066
without actually doing
a test drive.

985
00:43:58,100 --> 00:43:59,200
And these are things

986
00:43:59,233 --> 00:44:02,266
that involve negotiation
with other drivers.

987
00:44:02,300 --> 00:44:04,766
Because in order to check
if you are negotiating

988
00:44:04,800 --> 00:44:08,166
normally and properly
with other drivers,

989
00:44:08,200 --> 00:44:09,166
you need other drivers.

990
00:44:09,200 --> 00:44:11,800
NARRATOR:
Today, Mobileye engineers

991
00:44:11,833 --> 00:44:14,266
are testing a part
of their planning software

992
00:44:14,300 --> 00:44:17,133
that handles merging in traffic.

993
00:44:17,166 --> 00:44:19,400
So, we merged fine,

994
00:44:19,433 --> 00:44:21,533
but at the last point,
we're a bit slow.

995
00:44:21,566 --> 00:44:24,100
Yes, yes.

996
00:44:24,133 --> 00:44:25,366
Merging into traffic,

997
00:44:25,400 --> 00:44:27,633
this multi-agent game
that we are playing,

998
00:44:27,666 --> 00:44:30,200
requires
sophisticated negotiation.

999
00:44:30,233 --> 00:44:31,933
You are negotiating.

1000
00:44:31,966 --> 00:44:34,566
Your motion signals
to other road users your intent.

1001
00:44:34,600 --> 00:44:36,166
So you are negotiating.

1002
00:44:36,200 --> 00:44:39,833
And this negotiation
requires skills.

1003
00:44:39,866 --> 00:44:42,200
And those skills don't come
naturally.

1004
00:44:42,233 --> 00:44:43,733
Those skills need to be trained.

1005
00:44:43,766 --> 00:44:46,500
You need to now change two lanes

1006
00:44:46,533 --> 00:44:49,666
because the exit is a
few hundred meters, uh, from us.

1007
00:44:49,700 --> 00:44:54,066
NARRATOR:
The software is trained to be
assertive when it has to be.

1008
00:44:54,100 --> 00:44:57,400
Here, the car signals
its intention to exit

1009
00:44:57,433 --> 00:45:00,700
by speeding up so it can merge.

1010
00:45:00,733 --> 00:45:02,366
SHASHUA:
And you saw
that we changed two lanes

1011
00:45:02,400 --> 00:45:04,566
without obstructing
the flow of traffic

1012
00:45:04,600 --> 00:45:06,800
and, and there are
many vehicles here.

1013
00:45:06,833 --> 00:45:10,100
You need to provide agility
that is as good as humans

1014
00:45:10,133 --> 00:45:11,500
if you want to be safe.

1015
00:45:11,533 --> 00:45:13,366
NARRATOR:
The skill gap

1016
00:45:13,400 --> 00:45:14,900
between autonomous systems
and human drivers

1017
00:45:14,933 --> 00:45:16,500
is narrowing.

1018
00:45:16,533 --> 00:45:17,900
And to teach self-driving cars

1019
00:45:17,933 --> 00:45:21,100
to be even better than humans,

1020
00:45:21,133 --> 00:45:23,200
some engineers
are exploiting the things

1021
00:45:23,233 --> 00:45:26,333
that computers can already
do extremely well.

1022
00:45:26,366 --> 00:45:27,933
GERDES (on radio):
Okay, you may begin.

1023
00:45:27,966 --> 00:45:32,300
NARRATOR:
Like control machinery
with extraordinary precision.

1024
00:45:32,333 --> 00:45:34,433
JONATHAN GOH:
Autonomous mode
in three, two, one.

1025
00:45:38,933 --> 00:45:43,166
NARRATOR:
This car is drifting,
a kind of controlled skid.

1026
00:45:43,200 --> 00:45:45,933
And what makes
this feat possible

1027
00:45:45,966 --> 00:45:50,133
is a computer programmed
to exploit the laws of physics.

1028
00:45:50,166 --> 00:45:54,966
Here at Thunderhill Raceway
in California,

1029
00:45:55,000 --> 00:45:56,833
a team from Stanford University
is developing

1030
00:45:56,866 --> 00:45:58,433
self-driving software

1031
00:45:58,466 --> 00:46:02,233
that can take evasive action
to escape danger.

1032
00:46:02,266 --> 00:46:04,100
(cheer)

1033
00:46:04,133 --> 00:46:05,766
GERDES:
In an emergency situation,

1034
00:46:05,800 --> 00:46:08,933
you want to be able to use

1035
00:46:08,966 --> 00:46:11,000
all of the capabilities
of the tires

1036
00:46:11,033 --> 00:46:14,466
to do anything that's required
to avoid that collision.

1037
00:46:14,500 --> 00:46:17,766
Our automated vehicles
are able to put the vehicle

1038
00:46:17,800 --> 00:46:21,066
into a very heavy swerve

1039
00:46:21,100 --> 00:46:23,900
when that is the best choice
for how to avoid the collision.

1040
00:46:23,933 --> 00:46:27,733
NARRATOR:
The Stanford team applies
what they've learned

1041
00:46:27,766 --> 00:46:29,333
from the undisputed masters

1042
00:46:29,366 --> 00:46:31,733
of pushing car performance
to the limit:

1043
00:46:31,766 --> 00:46:34,733
race car drivers.

1044
00:46:34,766 --> 00:46:38,566
GERDES:
Race car drivers are always
pushing up to the limits.

1045
00:46:38,600 --> 00:46:42,433
But they're trying to avoid
accidents when they do that.

1046
00:46:42,466 --> 00:46:45,533
So, for instance, race car
drivers are trying to use

1047
00:46:45,566 --> 00:46:46,933
all the friction
between the tire and the road

1048
00:46:46,966 --> 00:46:48,366
to be fast.

1049
00:46:48,400 --> 00:46:50,666
We want to use all the friction
between the tire and the road

1050
00:46:50,700 --> 00:46:52,033
to be safe.

1051
00:46:52,066 --> 00:46:56,233
NARRATOR:
While it might take
a race car driver years

1052
00:46:56,266 --> 00:46:59,166
to perfect skills like these,

1053
00:46:59,200 --> 00:47:00,733
once the software masters them,

1054
00:47:00,766 --> 00:47:04,066
a download could pass them on
to a whole fleet of cars

1055
00:47:04,100 --> 00:47:05,933
in just minutes.

1056
00:47:05,966 --> 00:47:07,466
GERDES:
We feel this is

1057
00:47:07,500 --> 00:47:09,166
a fundamental building block
of any type

1058
00:47:09,200 --> 00:47:11,100
of automated vehicle
that you would want to develop.

1059
00:47:11,133 --> 00:47:14,666
The vehicle should be able
to use all of its capabilities

1060
00:47:14,700 --> 00:47:16,466
to move out of harm's way.

1061
00:47:16,500 --> 00:47:19,266
(on radio):
You guys rock.

1062
00:47:22,066 --> 00:47:24,333
NARRATOR:
So, despite all the obstacles,

1063
00:47:24,366 --> 00:47:27,233
many engineers think
they're closing in on the prize:

1064
00:47:27,266 --> 00:47:30,900
self-driving cars safe enough
to trust.

1065
00:47:30,933 --> 00:47:34,266
Cars that could proliferate
very rapidly.

1066
00:47:34,300 --> 00:47:35,900
If and when that happens,

1067
00:47:35,933 --> 00:47:39,933
they will share the road
with human drivers.

1068
00:47:39,966 --> 00:47:41,933
How will that work?

1069
00:47:41,966 --> 00:47:44,166
Will we all get along?

1070
00:47:44,200 --> 00:47:46,533
In Michael Fleming's
test drives around the country,

1071
00:47:46,566 --> 00:47:48,400
he sees a clash brewing.

1072
00:47:48,433 --> 00:47:51,566
His company's autonomous car

1073
00:47:51,600 --> 00:47:54,400
is named Asimov,
after Isaac Asimov,

1074
00:47:54,433 --> 00:47:58,433
famed for his
science fiction books on robots.

1075
00:47:58,466 --> 00:48:01,133
FLEMING:
Asimov has a clear-cut rule book

1076
00:48:01,166 --> 00:48:03,233
and is very consistent.

1077
00:48:03,266 --> 00:48:05,733
But oftentimes,
when we drive safely

1078
00:48:05,766 --> 00:48:07,700
and follow the letter
of the law,

1079
00:48:07,733 --> 00:48:09,966
we get honked at.

1080
00:48:10,000 --> 00:48:12,000
And do you know
who we get honked at by?

1081
00:48:12,033 --> 00:48:14,233
The aggressive drivers,
the rule breakers.

1082
00:48:14,266 --> 00:48:19,000
NARRATOR:
Fleming and his team
analyze the data

1083
00:48:19,033 --> 00:48:20,700
from the car's
strange encounters

1084
00:48:20,733 --> 00:48:24,033
and use it to improve
their software.

1085
00:48:24,066 --> 00:48:26,600
Whether it's a careless
pedestrian...

1086
00:48:26,633 --> 00:48:29,500
FLEMING:
This lady just steps out
in the middle of the road.

1087
00:48:29,533 --> 00:48:33,133
And if you notice,
she doesn't even turn her head.

1088
00:48:33,166 --> 00:48:35,333
NARRATOR:
Or a wrong-way car.

1089
00:48:35,366 --> 00:48:39,366
FLEMING:
We were driving down a one-way
street with multiple lanes.

1090
00:48:39,400 --> 00:48:41,166
And we see this white truck

1091
00:48:41,200 --> 00:48:45,433
driving the wrong way
down a one-way road.

1092
00:48:45,466 --> 00:48:47,766
So clearly we have
a rule breaker here,

1093
00:48:47,800 --> 00:48:49,666
doing something
that they shouldn't do.

1094
00:48:49,700 --> 00:48:54,166
You know, the development
of self-driving technology

1095
00:48:54,200 --> 00:48:56,533
would be pretty simple

1096
00:48:56,566 --> 00:48:57,966
if everyone just followed
the rules of the road.

1097
00:48:58,000 --> 00:49:01,100
If everyone came to a stop
at a stop sign.

1098
00:49:01,133 --> 00:49:03,633
If everyone used crosswalks.

1099
00:49:03,666 --> 00:49:05,700
But the reality is, they don't.

1100
00:49:05,733 --> 00:49:08,566
They break rules all the time.

1101
00:49:08,600 --> 00:49:12,133
NARRATOR:
Conflicts between
rule-bound automated cars

1102
00:49:12,166 --> 00:49:14,033
and impatient humans

1103
00:49:14,066 --> 00:49:16,700
are just one potential problem.

1104
00:49:16,733 --> 00:49:21,633
If robotaxi services
become very cheap,

1105
00:49:21,666 --> 00:49:24,600
might traffic actually increase,
and pollution grow worse?

1106
00:49:24,633 --> 00:49:26,533
How many people

1107
00:49:26,566 --> 00:49:28,533
who now earn their living
from driving

1108
00:49:28,566 --> 00:49:31,066
might lose their jobs?

1109
00:49:31,100 --> 00:49:35,200
If millions of cars
are electronically connected,

1110
00:49:35,233 --> 00:49:37,033
what risks might that pose

1111
00:49:37,066 --> 00:49:41,100
to our privacy
and our security?

1112
00:49:41,133 --> 00:49:44,433
Finally, an ethical question:

1113
00:49:44,466 --> 00:49:47,633
Are we willing to accept
self-driving cars

1114
00:49:47,666 --> 00:49:49,166
that kill some people,

1115
00:49:49,200 --> 00:49:51,166
as long as they kill
fewer people

1116
00:49:51,200 --> 00:49:53,700
than human drivers do?

1117
00:49:55,266 --> 00:50:00,000
Despite these looming questions,
proponents say self-driving cars

1118
00:50:00,033 --> 00:50:05,400
could make transportation
both easier and safer.

1119
00:50:05,433 --> 00:50:07,066
I think new technology offers

1120
00:50:07,100 --> 00:50:10,500
the biggest safety tool
that we've had in 100 years.

1121
00:50:10,533 --> 00:50:12,100
That's the cultural
transformation that's coming.

1122
00:50:12,133 --> 00:50:15,100
NARRATOR:
But it is far from certain

1123
00:50:15,133 --> 00:50:20,166
that driverless cars will
ever deliver on that promise.

1124
00:50:20,200 --> 00:50:22,566
CUMMINGS:
I think it is hubris to believe

1125
00:50:22,600 --> 00:50:24,966
that driving
is such a simple task

1126
00:50:25,000 --> 00:50:28,133
that, since there's so much more
automation in the world,

1127
00:50:28,166 --> 00:50:30,366
how hard could this be?

1128
00:50:30,400 --> 00:50:33,600
I'm a big fan of where we're
going with this technology,

1129
00:50:33,633 --> 00:50:36,900
but I also work on a day-to-day
basis with this technology,

1130
00:50:36,933 --> 00:50:41,566
and it's just simply not ready
for public consumption

1131
00:50:41,600 --> 00:50:45,166
to any verifiable degree
of safety.

1132
00:50:45,200 --> 00:50:46,400
LEVINSON:
I think that we,

1133
00:50:46,433 --> 00:50:48,533
as developers in the industry,

1134
00:50:48,566 --> 00:50:50,166
need to earn the public's trust,

1135
00:50:50,200 --> 00:50:51,866
and not the other way around.

1136
00:50:51,900 --> 00:50:53,666
I think we need to be able
to demonstrate

1137
00:50:53,700 --> 00:50:57,100
why our system is, in fact,
safer than human drivers.

1138
00:50:57,133 --> 00:51:01,333
NARRATOR:
If self-driving cars eventually
do win public trust,

1139
00:51:01,366 --> 00:51:05,166
their adoption
may be less of a revolution

1140
00:51:05,200 --> 00:51:08,833
than a slow evolution.

1141
00:51:08,866 --> 00:51:10,866
In my opinion,
the safe solutions today

1142
00:51:10,900 --> 00:51:13,933
work at low speeds
in low-complexity environments.

1143
00:51:13,966 --> 00:51:17,800
So this includes driving
on private roads,

1144
00:51:17,833 --> 00:51:22,500
on campuses, retirement
communities, airports.

1145
00:51:22,533 --> 00:51:24,733
But we do not have solutions

1146
00:51:24,766 --> 00:51:28,100
that work in general
at high speeds,

1147
00:51:28,133 --> 00:51:31,300
in congestion,

1148
00:51:31,333 --> 00:51:34,566
and in really difficult
road conditions.

1149
00:51:34,600 --> 00:51:36,400
Having a fully
autonomous vehicle

1150
00:51:36,433 --> 00:51:40,900
being able to take you
anywhere, anytime

1151
00:51:40,933 --> 00:51:42,433
is very, very far in the future.

1152
00:51:42,466 --> 00:51:44,433
In fact, I don't have
even a guess

1153
00:51:44,466 --> 00:51:46,633
as to how far in the future
that will be.

1154
00:51:46,666 --> 00:51:49,100
SHLADOVER:
It's not like a mobile phone
app,

1155
00:51:49,133 --> 00:51:51,400
where, you know, if the
mobile phone app doesn't work

1156
00:51:51,433 --> 00:51:52,566
ten percent of the time,

1157
00:51:52,600 --> 00:51:53,733
big deal.

1158
00:51:53,766 --> 00:51:55,633
This has got to work
all the time.

1159
00:51:55,666 --> 00:51:58,566
There's a pot of gold out there

1160
00:51:58,600 --> 00:52:00,300
at the end of the rainbow

1161
00:52:00,333 --> 00:52:03,433
for those who can actually
get this to work.

1162
00:52:03,466 --> 00:52:07,666
Now, the challenge is
how to get it to work safely.

1163
00:52:09,333 --> 00:52:13,466
♪

1164
00:52:16,833 --> 00:52:19,200
Major funding for "NOVA"
is provided by the following:

1165
00:52:52,800 --> 00:52:55,000
To order this "NOVA" program
on DVD,

1166
00:52:55,033 --> 00:53:00,266
visit ShopPBS
or call 1-800-PLAY-PBS.

1167
00:53:00,300 --> 00:53:04,433
This program is also available
on Amazon Prime Video.

1168
00:53:04,466 --> 00:53:08,500
♪