Dawn of the Driverless Car Horizon


Dawn of the Driverless Car

Similar Content

Browse content similar to Dawn of the Driverless Car. Check below for episodes and series from the same categories and more!

Transcript


LineFromTo

The motorcar has shrunk the world,

0:00:030:00:06

increased personal freedom,

0:00:060:00:08

and in so many ways, expanded our horizons.

0:00:080:00:11

But there's a flip side.

0:00:120:00:14

Cars have destroyed our environment,

0:00:140:00:16

poisoned the air we breathe,

0:00:160:00:18

and killed us in far more straightforward ways.

0:00:180:00:22

There's about a million deaths every year in this world due to traffic

0:00:230:00:27

accidents and I find this just

0:00:270:00:29

utterly unacceptable in the 21st century.

0:00:290:00:32

But all that's going to change.

0:00:320:00:35

Soon, we'll be in a position to have our automotive cake and eat it.

0:00:350:00:39

Self-driving cars are going to have a huge impact on society.

0:00:400:00:44

They'll be able to navigate through complex intersections

0:00:440:00:47

with no collisions.

0:00:470:00:48

If I could just sit back and read a book, listen to music,

0:00:480:00:51

catch up on some sleep, that would be great.

0:00:510:00:53

Once your car can come from round the corner,

0:00:530:00:56

we can start opening up some of these sort of residential areas and

0:00:560:00:59

they'll feel much more sociable again.

0:00:590:01:01

This is a world where cars will drive themselves,

0:01:010:01:05

a world where we are simply passengers,

0:01:050:01:07

ferried about by wholesome, green,

0:01:070:01:09

compassionate technology which will never, ever go wrong.

0:01:090:01:13

And it's almost here.

0:01:130:01:15

I can press this button.

0:01:150:01:17

I don't even have a control that I can grab.

0:01:170:01:20

A fully autonomous vehicle in commercial operation in 2021.

0:01:200:01:26

But cars that can run errands for us by themselves could quickly clog up

0:01:270:01:31

our streets and ruin livelihoods, too.

0:01:310:01:34

From taxi drivers to truck drivers,

0:01:360:01:39

lots of people have a job as a driver.

0:01:390:01:43

A lot of effort has been put into selling us the driverless dream.

0:01:430:01:48

Now it's almost upon us,

0:01:480:01:50

could we actually be sleepwalking into a nightmare?

0:01:500:01:53

What happens when it eventually encounters a no-win scenario,

0:01:540:01:57

when it actually has to have an accident?

0:01:570:01:59

We love cars.

0:02:130:02:14

We love owning them,

0:02:140:02:17

we love driving them,

0:02:170:02:20

and learning to drive a car is a rite of passage.

0:02:200:02:23

Handbrake off. Handbrake.

0:02:230:02:25

-Oop.

-Handbrake.

0:02:250:02:27

You've got to be able to drive or else you're really not a real adult.

0:02:270:02:30

But learning to mirror, signal and manoeuvre is on its way out,

0:02:320:02:37

because shadowy backroom technologists at places like Google,

0:02:370:02:40

Intel and even Facebook

0:02:400:02:42

are hell-bent on getting rid of drivers altogether.

0:02:420:02:45

We're building systems to enable cars to drive themselves.

0:02:460:02:50

Our lovely cars are on borrowed time.

0:02:540:02:57

There's even a plan drawn up by the Society of Automotive Engineers.

0:02:570:03:01

A road map, if you will.

0:03:010:03:03

It starts with level zero.

0:03:040:03:07

Level zero is where the human driver has control of everything.

0:03:070:03:11

The plan then moves through various

0:03:120:03:14

levels of automation and driver-assist

0:03:140:03:17

technologies and ends up at level five.

0:03:170:03:20

Level five is where humans are simply passengers.

0:03:210:03:24

At this point, the word "driving"

0:03:250:03:28

reverts to being something to do with livestock.

0:03:280:03:31

If this sounds a bit like a science fiction writer's pipe dream,

0:03:330:03:36

then remember, an awful lot of effort and, crucially, money

0:03:360:03:40

is currently going into it.

0:03:400:03:42

Google is investing 30 million annually to driverless,

0:03:430:03:47

and Intel recently paid 15 billion

0:03:470:03:50

for Israeli driverless tech company Mobileye.

0:03:500:03:53

From niche robot race cars...

0:03:550:03:58

Look, it's being steered by nobody!

0:03:580:04:00

..to mass manufacturers,

0:04:010:04:04

driverless is where it's at.

0:04:040:04:06

I think the promise of a driverless future where cars are available to

0:04:070:04:12

everyone and they can do all the

0:04:120:04:14

hard work and ease congestion and ease pollution is very exciting,

0:04:140:04:19

but I am a bit doubtful as to

0:04:190:04:20

whether that's how it will actually play out.

0:04:200:04:23

Autonomy is probably the biggest

0:04:230:04:25

thing that's being talked about in the automotive industry.

0:04:250:04:28

People that are less able will be able to get around because they

0:04:280:04:31

won't need to be driving. But obviously,

0:04:310:04:33

there's so much legislation to get in place to get to the point where

0:04:330:04:36

-this is actually viable.

-I do think people want...

0:04:360:04:39

In principle, they want to be able to move around with less effort.

0:04:390:04:42

I think everybody wants that.

0:04:420:04:44

I'm not sure that's what Ford is talking about when it says it's

0:04:440:04:46

going to have a driverless car by 2020, or whatever it is.

0:04:460:04:49

It sounds exciting,

0:04:500:04:52

but I think it's...

0:04:520:04:55

I want to make sure it works, and that's the biggest challenge.

0:04:550:04:58

Is it going to do what it's supposed to do?

0:04:580:05:01

Will I trust it?

0:05:010:05:03

Whatever eventually emerges onto the roads of tomorrow,

0:05:030:05:06

the future looks bleak for the intimate relationship we currently

0:05:060:05:09

enjoy with our fine, four-fendered friends.

0:05:090:05:12

This is a momentous day in Harrison's life.

0:05:200:05:24

Today he is having his first driving lesson.

0:05:240:05:27

I learned to drive when I was 17.

0:05:280:05:31

I was quite nervous.

0:05:310:05:33

I loved learning to drive. I'd been

0:05:330:05:34

looking forward to it for quite a long time.

0:05:340:05:36

I was failed pretty much about

0:05:360:05:38

15 minutes in for failing to give way to someone.

0:05:380:05:40

Harrison has never attempted to drive before,

0:05:400:05:43

and now he's nailed the

0:05:430:05:44

all-important selfie with his new best friend,

0:05:440:05:47

there's a lot to think about.

0:05:470:05:49

Push that all the way down?

0:05:490:05:50

Yeah, press the brake all the way down and press the start button.

0:05:500:05:53

Look all around the car like this.

0:05:530:05:55

Just pop your signal on, press the button in,

0:05:550:05:57

pull it up and ease up off your clutch, slowly.

0:05:570:05:59

-That's it.

-OK.

-And we're just going to go off like that.

0:05:590:06:02

-OK.

-Have you got the idea?

-Yeah.

0:06:020:06:04

In a few short minutes, however,

0:06:040:06:06

Harrison is transformed from ordinary mortal to driver.

0:06:060:06:11

Exciting, isn't it? Yo, let's go!

0:06:110:06:13

-Let's go.

-Don't worry, don't worry.

0:06:130:06:16

-That's fine.

-OK, so he's not the finished article,

0:06:160:06:18

but in a couple of months,

0:06:180:06:20

he'll hopefully have passed his driving test and

0:06:200:06:22

will think nothing more of driving than he currently does of walking.

0:06:220:06:26

There you go. You're off.

0:06:260:06:28

You'll be driving back home in no time.

0:06:280:06:31

In simple terms,

0:06:310:06:32

Harrison has sensed the world around him and reacted appropriately.

0:06:320:06:36

It's what he's been doing all his life.

0:06:360:06:38

-Well done.

-Only today, he's learning to do that via a machine.

0:06:390:06:45

Sensing our environment is something most of us take wholly for granted.

0:06:450:06:49

It's part of being human, something we're good at.

0:06:490:06:53

But little by little, cars have been getting in on the act, too.

0:06:530:06:56

Parking sensors, lane sensors, cruise control,

0:07:030:07:08

adaptive cruise control,

0:07:080:07:10

automatic headlights and windscreen wipers have all emerged under the

0:07:100:07:14

banner of driver-assist technologies.

0:07:140:07:16

But in reality, they are the first,

0:07:180:07:21

small steps on the road to full autonomy, and that's official.

0:07:210:07:25

It's the first stage on the driverless masterplan.

0:07:250:07:29

Level one.

0:07:310:07:34

At level one, the vehicle can take control of individual functions like

0:07:340:07:38

acceleration or braking.

0:07:380:07:40

Effectively, cars can actually drive

0:07:410:07:44

themselves to a small degree already.

0:07:440:07:46

Electronic stability control or ABS is one of the systems that's been

0:07:470:07:52

put in that stops me from skidding and crashing.

0:07:520:07:55

Adaptive cruise control brakes and speeds up for you,

0:07:550:07:58

depending on what the car in front

0:07:580:08:00

is doing, and that is just so great for long journeys.

0:08:000:08:02

Like, I drove to Wales on the M4 the other week and it's a

0:08:020:08:05

really long journey and it was just effortless because of that.

0:08:050:08:09

Many of these automotive technologies

0:08:090:08:11

have their origins at the racetrack.

0:08:110:08:13

And even though car racing is about as driver-focused

0:08:170:08:20

as it's possible to be...

0:08:200:08:22

..there is some common ground

0:08:240:08:26

between Formula 1 and the driverless future.

0:08:260:08:28

My name is Teena Gade, and I work here at Sahara Force India Formula 1

0:08:320:08:35

team as a vehicle science engineer.

0:08:350:08:37

The best place normally to develop a car is actually to take it to a

0:08:380:08:41

track and test it. In reality,

0:08:410:08:42

we're restricted by the regulations so we're only allowed a fixed number

0:08:420:08:45

of test sessions every year.

0:08:450:08:47

What that means is we end up having to do quite a lot of it in the

0:08:470:08:50

virtual world. We model the car, the tyres,

0:08:500:08:52

the aerodynamics and the tracks and

0:08:520:08:54

we bring it all together in a driving

0:08:540:08:56

simulator for a driver to drive around and it tells us what the

0:08:560:08:58

performance is going to be.

0:08:580:09:00

One of the most interesting things for me about the concept of

0:09:000:09:03

driverless cars is if you take what I do on a daily basis

0:09:030:09:05

but we really don't understand,

0:09:050:09:07

the bit that is difficult is the driver element, because for all the

0:09:070:09:10

computational power we have to model, for example,

0:09:100:09:13

the aerodynamics or the tyres or the track,

0:09:130:09:15

the things that go on in the human brain are incredibly complicated and

0:09:150:09:18

processed unbelievably quickly.

0:09:180:09:20

So from an engineering perspective,

0:09:200:09:23

if we could have a completely driverless car it could perform the

0:09:230:09:25

same task absolutely consistently, the same from one lap to the next,

0:09:250:09:29

every time. We would get much cleaner data and actually,

0:09:290:09:32

we potentially stand to make better connections

0:09:320:09:34

as a result and learn more.

0:09:340:09:36

Today, Teena is off to see another race team,

0:09:380:09:41

one that have taken driverless technology a stage further.

0:09:410:09:45

I actually think I want to go left here.

0:09:460:09:48

No, I was completely wrong. I can go that way.

0:09:480:09:51

Roborace want to compete in Formula E,

0:09:540:09:58

but unlike other electric race teams,

0:09:580:10:00

they have decided to abandon drivers altogether.

0:10:000:10:03

These sleek racing robots will battle it out against each other and

0:10:050:10:09

eventually against human competition.

0:10:090:10:11

But for the time being, they're testing out the concept with this,

0:10:120:10:16

their development robot, DevBot.

0:10:160:10:19

We can see inside here,

0:10:210:10:23

it actually looks quite a lot like a conventional car cabin.

0:10:230:10:26

Yeah, we've got a rack that sits behind the driver.

0:10:260:10:29

What we've done is we've taken human capabilities and we're putting

0:10:290:10:32

them in silicon and software.

0:10:320:10:34

Here at Silverstone, DevBot knows

0:10:360:10:39

the circuit so well that it's capable of

0:10:390:10:41

giving high-speed tours of the racetrack to the humans it's aiming

0:10:410:10:45

-to defeat.

-I'm not actually the best passenger at the best of times,

0:10:450:10:49

and so it's going to be quite strange.

0:10:490:10:52

Hold the handbrake.

0:10:520:10:54

-I would drive like that.

-OK.

0:10:540:10:56

So this is it.

0:10:580:11:00

My life in the hands of some software.

0:11:000:11:02

The blue light's gone on at the front.

0:11:020:11:04

I think we're ready to go.

0:11:040:11:05

I am really very nervous.

0:11:050:11:07

It's terrifying, the first time you get in a car and you're not touching

0:11:140:11:18

any of the controls.

0:11:180:11:19

Once you've spent some time in it,

0:11:390:11:40

you get that feeling of total confidence,

0:11:400:11:43

and you realise that the machine is

0:11:430:11:44

far better than any human could ever be.

0:11:440:11:46

Oh, wow!

0:11:460:11:48

That is hard on the brakes.

0:11:480:11:50

Can I have another go? That's actually quite good!

0:11:500:11:53

One of the big appeals about Formula 1 and motorsport in general is the

0:11:530:11:56

personalities, the drivers themselves.

0:11:560:11:59

There is also a huge following for teams.

0:11:590:12:02

It's quite good in there, it's quite good, yeah!

0:12:020:12:04

If you look at Ferrari, they're one of the biggest in the world.

0:12:040:12:07

And I think what Roborace will allow is actually if people create that

0:12:070:12:10

following for the team rather than for the driver themselves.

0:12:100:12:14

Impressive as careering driverless

0:12:150:12:18

around the track at around 200kph is,

0:12:180:12:21

there's much more to racing than just pure speed.

0:12:210:12:25

A driverless race car will have a lot on its mind.

0:12:250:12:28

What we're really starting to look at is that judgment layer.

0:12:300:12:33

"What speed should I be entering this corner?" -

0:12:330:12:36

is the critical thing.

0:12:360:12:37

The path that I should be following.

0:12:370:12:40

And then moving up into sort of the tactical decision-making layer in

0:12:400:12:43

terms of - "Am I going to overtake or am I going to actually save some

0:12:430:12:46

"battery because I want to attack in about five laps' time?"

0:12:460:12:50

So what you're basically saying is

0:12:500:12:51

that actually the car can perform the driving function as a human can?

0:12:510:12:55

Yes, that's exactly right, yeah.

0:12:550:12:57

It is Roborace's plan to build such

0:12:580:13:01

a clever driverless car that the human

0:13:010:13:03

opposition will be ground into the dust.

0:13:030:13:06

But so far, the DevBots are only racing each other.

0:13:060:13:09

With mixed results.

0:13:090:13:11

And this is the problem.

0:13:140:13:16

Robots, even quite advanced ones

0:13:160:13:19

like the ones that run DevBot, are, well, limited.

0:13:190:13:24

Here is a representative selection.

0:13:280:13:30

Looking at these brand ambassadors,

0:13:320:13:34

the future for autonomous cars doesn't look over-rosy.

0:13:340:13:37

If they stand any chance of gaining our trust,

0:13:380:13:42

they're going to have to deliver a lot more than these.

0:13:420:13:45

And DevBot.

0:13:460:13:48

Yeah, so we're turning left. Ease up very slowly off the clutch.

0:13:530:13:56

And start to steer towards me.

0:13:560:13:59

To successfully avoid crashing his car,

0:13:590:14:01

Harrison will have to be able to perform a bewildering array of tasks

0:14:010:14:05

simultaneously in real time.

0:14:050:14:08

You've gone in the wrong gear, don't worry.

0:14:080:14:10

-Oh, sorry.

-I've helped you out. Don't worry.

0:14:100:14:12

That's part of being a learner.

0:14:120:14:14

-OK.

-Harrison will have to be able to recognise, categorise,

0:14:140:14:17

and accurately predict the likely

0:14:170:14:19

future actions of anything he sees while driving.

0:14:190:14:22

Check this mirror

0:14:220:14:24

and put your left signal on, which is downwards.

0:14:240:14:27

In the light of those

0:14:270:14:28

near instantaneously acquired pieces of information,

0:14:280:14:31

he must cause the car to safely accelerate, change direction,

0:14:310:14:34

slow down or stop, or all or none of the above.

0:14:340:14:38

Hang on, I'm just going to stop you there.

0:14:380:14:40

We just had a car coming rather quickly.

0:14:400:14:42

And bear in mind that this is a constantly updated stream of data to

0:14:460:14:51

be interpreted and acted upon by Harrison,

0:14:510:14:54

millisecond by millisecond,

0:14:540:14:55

for as long as he's behind the wheel.

0:14:550:14:58

We just had a narrow escape with that little caravan there.

0:14:580:15:01

That, in a nutshell, is what will be required of a self-driving car.

0:15:020:15:07

It too will need to understand and interact with its environment.

0:15:070:15:12

But let's not get ahead of ourselves.

0:15:120:15:14

-We're not going down that busy one, don't worry.

-Yeah, good!

0:15:140:15:17

We're only just getting to level two, after all.

0:15:180:15:21

Level two.

0:15:210:15:23

Level two is the point on the road to autonomy at which the would-be

0:15:230:15:27

driverless car can control two things at the same time,

0:15:270:15:31

like steering AND braking.

0:15:310:15:33

It's a small but significant step.

0:15:350:15:38

But before we send all our current cars to the crusher,

0:15:380:15:41

it's worth remembering that some of them can multitask already.

0:15:410:15:45

Automatic parking I think people are very happy to use.

0:15:460:15:50

It scans for a car parking space,

0:15:500:15:52

and then you just have to do the

0:15:520:15:54

gas and brake while it does the steering for you.

0:15:540:15:56

All these tricks are all well and good,

0:15:560:15:59

and of course really very clever.

0:15:590:16:01

But to be fair, they're not exactly

0:16:010:16:03

difficult for human drivers to pull off.

0:16:030:16:06

It makes you wonder why we're

0:16:060:16:08

bothering with driverless cars at all.

0:16:080:16:10

The main attractive feature of

0:16:100:16:12

driverless cars is the promise that they will save lives.

0:16:120:16:17

I don't think humans are that great at driving.

0:16:170:16:19

You only have to look at the statistics to see that

0:16:190:16:22

there is a need for improvement there.

0:16:220:16:25

So if automating the driver can reduce those deaths,

0:16:250:16:30

then it's definitely a desirable thing.

0:16:300:16:32

There are systems that can detect whether you fall asleep or not.

0:16:330:16:36

And if you don't wake up, then the car will perform an emergency stop.

0:16:360:16:40

And some really cool stuff has been happening in collision avoidance,

0:16:400:16:43

so the car taking over when it

0:16:430:16:45

thinks you're going to have an accident

0:16:450:16:48

and you haven't reacted quickly enough.

0:16:480:16:50

The number of accidents that's saved alone over the last couple of years

0:16:500:16:54

is phenomenal.

0:16:540:16:55

The promise of a safer-than-human driverless car has been around for

0:16:570:17:01

almost as long as the car itself.

0:17:010:17:03

We trust our lives in machines all the time.

0:17:070:17:10

Every time you get on a commercial flight,

0:17:120:17:14

you're not being flown by a human any more.

0:17:140:17:16

And that's good, because you're safer this way.

0:17:160:17:18

Sebastian Thrun has spent most of

0:17:180:17:20

his professional life trying to bring

0:17:200:17:23

that airline level of safety to our cars.

0:17:230:17:25

I got involved because I had a traumatic event as a teenager.

0:17:260:17:30

My best friend died in a traffic

0:17:300:17:31

accident from one moment to the next.

0:17:310:17:34

And I found his death kind of a little bit ridiculous.

0:17:340:17:37

I think we don't talk about that much,

0:17:370:17:38

but there's about 1 million or 1.2 million

0:17:380:17:41

deaths every year in this world to traffic accidents.

0:17:410:17:44

And I find this just utterly unacceptable in the 21st century.

0:17:440:17:47

So I wanted to fix that.

0:17:470:17:49

Given that 90% of traffic accidents are due to

0:17:490:17:52

human error, Sebastian decided that the best and safest course of action

0:17:520:17:56

would be to dispense with the driver altogether.

0:17:560:17:59

In 2005, he had a breakthrough.

0:18:030:18:05

My team at Stanford built the car that won the DARPA Grand Challenge,

0:18:070:18:11

a desert race for a car that could drive itself.

0:18:110:18:14

The car crossed more than 200 kilometres of desert in a little

0:18:160:18:20

over six hours, with no human driver and no human intervention.

0:18:200:18:24

Sebastian was delighted.

0:18:260:18:28

The DARPA Grand Challenge winner and DevBot show a glimpse of the future.

0:18:320:18:37

But the trouble is that the world of driving doesn't usually exist on a

0:18:370:18:40

racetrack of known shape, size and camber

0:18:400:18:44

or in a relatively benign desert,

0:18:440:18:46

where the worst that could happen is cactus damage.

0:18:460:18:48

Real driving happens in the temporarily reversed Croydon one-way

0:18:500:18:54

system on a wet Thursday morning.

0:18:540:18:56

Formula 1 engineer Teena Gade has been impressed by the idea of robot

0:19:000:19:05

race cars. But like the rest of us,

0:19:050:19:07

she is keen to discover how

0:19:070:19:08

driverless cars might help with the real world.

0:19:080:19:13

Driving is pleasurable, say, for example,

0:19:130:19:15

you're driving on the west coast of California.

0:19:150:19:17

But if you do it for ten hours a week,

0:19:170:19:19

suddenly it's not that much fun any more.

0:19:190:19:22

The reality is that if there was a

0:19:220:19:24

train that went between where I lived and worked,

0:19:240:19:26

I'd probably take the train,

0:19:260:19:27

because it would mean I could be doing something more productive.

0:19:270:19:30

So if I had a machine that could do that for me,

0:19:300:19:33

then that would be a great thing.

0:19:330:19:35

That would buy me back ten hours a week.

0:19:350:19:37

That's a whole working day.

0:19:370:19:38

Today, Teena's off to visit a tech start-up company who want to help.

0:19:410:19:44

FIVE AI is the brainchild of Stan Boland,

0:19:460:19:50

who aims to turn this reasonably

0:19:500:19:52

priced electric car into the star of the driverless world.

0:19:520:19:56

So why would you call your company FIVE AI?

0:19:570:19:59

We're aiming for, ultimately,

0:19:590:20:01

the highest level of autonomy, which is level five,

0:20:010:20:04

a car that is utterly autonomous,

0:20:040:20:06

can drive anywhere without any kind of human intervention.

0:20:060:20:10

In fact, it hasn't even got provision for human intervention.

0:20:100:20:13

So we called the company FIVE AI.

0:20:130:20:14

Oh, wow, OK, that makes sense.

0:20:140:20:16

Urban driving creates the most challenges, actually,

0:20:160:20:19

for an autonomous car system.

0:20:190:20:21

You may have cyclists and pedestrians and cars,

0:20:210:20:25

complex buildings,

0:20:250:20:27

road markings that confuse, and people might come from any direction

0:20:270:20:30

and do almost anything in front of you, really.

0:20:300:20:32

Stan and the team are building their system from scratch, and today,

0:20:340:20:38

they're installing the sensors the

0:20:380:20:40

car will need to negotiate the world.

0:20:400:20:43

It needs to be able to see,

0:20:430:20:45

so we use cameras for that and we use light radar.

0:20:450:20:48

We also use ultrasound,

0:20:480:20:50

just like we would in our cars at

0:20:500:20:53

home to detect objects that are very, very close by.

0:20:530:20:55

But when you get to sort of solve problems like fog or

0:20:550:20:59

night-time, or rain or snow,

0:20:590:21:02

then it requires infrared cameras or high-sensitivity cameras,

0:21:020:21:05

so we should be able to sort of

0:21:050:21:07

drive along and just reconstruct the world as we go.

0:21:070:21:10

Just like humans do, really.

0:21:100:21:11

So the vision systems are actually just the tip of the iceberg.

0:21:110:21:14

What you're talking about is going on to then what's happening in the

0:21:140:21:17

-brains of this car?

-That's right,

0:21:170:21:19

it does require a huge amount of cognitive capability in the car to

0:21:190:21:22

be able to sort of sense the world and then take decisions about how we

0:21:220:21:26

control it. So that is a huge problem,

0:21:260:21:29

but one that we think we can actually solve.

0:21:290:21:33

To do that, Stan and the team will need to get their system through all

0:21:330:21:36

the levels of autonomy, including level three.

0:21:360:21:40

Level three.

0:21:430:21:45

At level three, the masterplan

0:21:460:21:48

states that safety-critical functions can

0:21:480:21:51

be completely assigned to the vehicle under certain driving or

0:21:510:21:54

environmental conditions...

0:21:540:21:55

..but that a supervising human driver

0:21:560:21:59

must be present to take over in emergencies.

0:21:590:22:02

What that means in reality is that if the driving and environmental

0:22:030:22:07

conditions allow, you can shout at

0:22:070:22:09

the children face-to-face while the car drives itself.

0:22:090:22:13

Now, this is much more like it, and what's more,

0:22:130:22:17

it's almost available now.

0:22:170:22:19

And this is what it might look like in reality.

0:22:200:22:23

Or at least, what it looks like in the mind of whichever creative Volvo

0:22:230:22:27

hired to create this.

0:22:270:22:29

Once the two green bars in the centre meet,

0:22:290:22:32

the paddle lights shift to green and the autopilot confirms that the

0:22:320:22:36

driving and the supervision is delegated to the car.

0:22:360:22:39

And this is what this slightly

0:22:420:22:43

nervous-looking man from Nissan unleashed

0:22:430:22:46

on some carefully selected parts of east London earlier this year.

0:22:460:22:50

This is a real car and these are real streets.

0:22:510:22:54

And those are real hands,

0:22:540:22:56

poised to grab the autonomous steering wheel at a moment's notice.

0:22:560:23:00

Being in an autonomous car is

0:23:010:23:03

obviously a weird feeling because you're doing nothing.

0:23:030:23:07

It's quite unnerving the first time you do it,

0:23:070:23:09

certainly handing over control on a motorway, for example,

0:23:090:23:12

when it's obviously high speed.

0:23:120:23:14

But what's surprising is how quickly you become accustomed to the fact

0:23:140:23:17

that the car is driving itself.

0:23:170:23:20

Driverless cars are kept on the

0:23:220:23:24

straight and narrow using three things.

0:23:240:23:26

First, satellite navigation knows, more or less, where the road is.

0:23:260:23:31

Lidar, a spinning laser,

0:23:310:23:33

builds up an overview picture of the environment by recording its own

0:23:330:23:37

reflection, and radar does a similar thing for the short range.

0:23:370:23:40

It's this combination that Sebastian Thrun and his team used

0:23:420:23:45

to autonomously navigate through the Mojave Desert

0:23:450:23:48

and win the DARPA Challenge.

0:23:480:23:50

But for autonomously driving you home from a drink-and-drugs-fuelled

0:23:520:23:55

weekend with your crazy friends on the coast, it just can't cut it.

0:23:550:23:59

You still need a human, and that's a problem...

0:24:000:24:02

..one that's being addressed at Stanford University in California.

0:24:040:24:07

I'm going to get over this hill and then I'll talk to you.

0:24:100:24:15

I'm Wendy Ju and I work on research particularly around how people are

0:24:150:24:19

going to interact with automation.

0:24:190:24:21

So we're basically borrowing a little piece of the possible future

0:24:210:24:25

and then running experiments to see how people react there.

0:24:250:24:28

So right now, we're in a driving simulator.

0:24:280:24:31

In this particular set-up, we are trusting

0:24:310:24:33

this robot steering wheel.

0:24:330:24:35

I can press this button.

0:24:350:24:37

Now the car is going to switch to autonomy,

0:24:380:24:40

but the steering wheel handle actually goes back.

0:24:400:24:43

So if I'm a little distracted, daydreaming and I look down,

0:24:430:24:46

I know who's in charge,

0:24:460:24:47

not only because there's this little icon on the dashboard,

0:24:470:24:50

but I don't even have a control I can grab.

0:24:500:24:52

The problem for the driver, though,

0:24:530:24:55

is paying attention when the machine is in control.

0:24:550:24:59

Human minds tend to wander, get distracted, or simply shut down.

0:24:590:25:03

The idea of autonomous cars is really, really exciting.

0:25:040:25:08

So it's ironic that the reality of being in an autonomous car is that

0:25:080:25:13

being in an autonomous car is really, really boring.

0:25:130:25:16

It goes against intuition,

0:25:160:25:18

but what our research points to is how to keep people engaged and alert

0:25:180:25:22

and aware of what's going on on the road.

0:25:220:25:25

Because when we run an experiment,

0:25:250:25:27

people who are told just to

0:25:270:25:29

supervise the car are falling asleep.

0:25:290:25:31

Just the task of supervising the car

0:25:310:25:34

is not engaging enough to keep people awake.

0:25:340:25:37

And this is the Achilles heel of level three autonomy.

0:25:370:25:40

There needs to be a human ready to avert disaster at all times.

0:25:400:25:45

So we have this camera here pointed at my face to tell if they're awake

0:25:450:25:50

or asleep, if they're alert or not.

0:25:500:25:53

I think that it's going to become important for the car to do active

0:25:530:25:56

intervention. Maybe talking to people more,

0:25:560:25:58

entertain them more if they look drowsy or asleep.

0:25:580:26:01

Ask them questions, you know,

0:26:010:26:02

"Do you think we should go this way or that way?"

0:26:020:26:05

You know, "What do you think of this thing over there?"

0:26:050:26:07

If only the humans inside had something to do.

0:26:070:26:10

Like steer.

0:26:100:26:11

Just a thought.

0:26:130:26:14

I think people want fully driverless cars that can do absolutely

0:26:160:26:20

everything without human input.

0:26:200:26:23

I'm less convinced that people will

0:26:240:26:26

be happy with a supposedly driverless

0:26:260:26:28

car that actually needs them to be

0:26:280:26:30

paying attention the whole time in case something goes wrong.

0:26:300:26:33

What happens when we get used to not driving for

0:26:330:26:36

long periods of time and then are asked to act quickly if something

0:26:360:26:40

happens and take control,

0:26:400:26:42

and our reactions are not fast enough to actually do that?

0:26:420:26:45

Will they be completely distracted reading a magazine and so not

0:26:450:26:49

realise they have to take control?

0:26:490:26:51

That's one of the biggest challenges.

0:26:510:26:53

Level three automation, then, is slightly pointless,

0:26:540:26:57

a kind of cleverer cruise control

0:26:570:27:00

with a bit of steering sometimes thrown in.

0:27:000:27:02

What we really need is something

0:27:020:27:04

that can take on everything a human driver can.

0:27:040:27:07

Something clever. Intelligent, even.

0:27:070:27:11

Luckily, help is at hand on the

0:27:140:27:16

other side of the San Francisco Bay Area.

0:27:160:27:18

The University of California at Berkeley is Stanford's groovier,

0:27:200:27:24

longer-haired, more radical cousin.

0:27:240:27:26

Scientists here have come up with this.

0:27:290:27:32

It may not look radical,

0:27:320:27:34

but what it can do may just deliver the driverless revolution we've all

0:27:340:27:37

been promised.

0:27:370:27:38

Brett is the Berkeley Robot for the Elimination of Tedious Tasks.

0:27:400:27:44

If you look here, we got Brett's toys.

0:27:450:27:47

Brett has learned to stack Lego blocks,

0:27:470:27:51

has learned to put caps onto bottles,

0:27:510:27:55

assembling pieces of this airplane.

0:27:550:27:57

Whilst this might seem slightly underwhelming,

0:27:570:28:00

the key here is the idea of learning.

0:28:000:28:04

Nobody has written Brett a bottle-top program or a block-stacking

0:28:040:28:08

algorithm. The robot has worked it out for himself,

0:28:080:28:13

something that until recently was the exclusive domain

0:28:130:28:16

of animals with sizeable brains.

0:28:160:28:18

I am Pieter Abbeel.

0:28:190:28:21

I'm a professor at UC Berkeley and I work on artificial intelligence for

0:28:210:28:24

robotics. Basically,

0:28:240:28:26

robots are mechanically very capable and can do a lot of things.

0:28:260:28:29

But in practice, they do very little for us.

0:28:290:28:31

What's holding them back is their lack of intelligence.

0:28:310:28:34

Or at least it was.

0:28:370:28:39

Artificial intelligence, also known as machine learning,

0:28:390:28:43

is allowing robots to acquire new skills for themselves.

0:28:430:28:47

Just like we do.

0:28:470:28:49

So imitation learning is the process of having a robot learn from

0:28:490:28:52

watching something done for it.

0:28:520:28:54

Then in the future, faced with a different situation,

0:28:540:28:57

as to tie a knot, we're going to see the robot

0:28:570:29:00

understand what to do in the new situation.

0:29:000:29:03

For this challenge,

0:29:030:29:04

the rope is relayed in roughly the same place and Brett is asked to tie

0:29:040:29:08

the knot again.

0:29:080:29:10

Right, let's see what Brett can do now.

0:29:100:29:12

Brett can't simply repeat the exact motions he's been shown because the

0:29:140:29:17

rope isn't in exactly the same place.

0:29:170:29:19

He'll have to adapt the principles

0:29:210:29:22

of the challenge to the reality of the rope's new position.

0:29:220:29:26

It's not just repeating the motions we gave.

0:29:270:29:30

It's looking at how the new

0:29:300:29:31

situation relates to the old situation and

0:29:310:29:34

then morphing the motions,

0:29:340:29:36

make them the right motions for the new situation.

0:29:360:29:38

Yeah, there, Brett did it.

0:29:380:29:40

Beautiful knot.

0:29:400:29:42

Beautiful indeed, but Brett's skills don't end at imitation.

0:29:420:29:46

He can work out how to perform a task from scratch, like a toddler.

0:29:480:29:52

A rather unattractive toddler with metal arms and an electronic brain.

0:29:530:29:58

But if he can learn, there's no reason why one day Brett,

0:29:580:30:02

or a robot like him, couldn't learn to drive.

0:30:020:30:04

But we're getting ahead of ourselves.

0:30:060:30:09

We'll specify the objective rather

0:30:090:30:11

than the strategy to achieve something.

0:30:110:30:13

So you don't have to demonstrate

0:30:130:30:14

everything multiple times for the robot to understand what to do.

0:30:140:30:17

You just give the robot an objective.

0:30:170:30:19

Then the robot just can go at it,

0:30:190:30:22

pretty much like a kid playing,

0:30:220:30:23

trying, trying, trying and, over time,

0:30:230:30:26

getting better and better.

0:30:260:30:27

This challenge requires Brett to get

0:30:290:30:31

the red cube into the box through one of the holes.

0:30:310:30:35

But he's not shown how.

0:30:350:30:36

He has no model of how his own arm works,

0:30:380:30:40

so initially the best he can do is random motion.

0:30:400:30:43

So right now...

0:30:430:30:44

As it refines that model and gets more data from its own attempts it

0:30:460:30:49

gets better and better at finding a

0:30:490:30:50

solution to get the block into the matching opening.

0:30:500:30:53

And actually, very quickly.

0:30:530:30:55

This took maybe one minute of learning

0:30:550:30:56

and it invented how to do that from scratch.

0:30:560:30:59

With imitation learning,

0:30:590:31:01

Brett simply adapts an action that

0:31:010:31:03

he's been shown, but machine learning,

0:31:030:31:06

artificial intelligence, allows him to solve a problem for himself.

0:31:060:31:10

The implications for his descendants and ours are far-reaching.

0:31:120:31:16

Maybe you want a robot to jump higher than a human has ever jumped.

0:31:180:31:21

Or you want to do something more precise than a human has ever done.

0:31:210:31:24

You can give it the objective and then it'll, on its own, try.

0:31:240:31:28

Not succeed initially, fail most of the time,

0:31:280:31:30

but over time get better and better and maybe exceed human performance.

0:31:300:31:34

When it comes to driverless cars,

0:31:360:31:38

exceeding human performance is essential.

0:31:380:31:41

After all, not many of us would want to be driven around by Brett.

0:31:430:31:46

Artificial intelligence is the

0:31:520:31:54

latest technological new kid on the block

0:31:540:31:57

and areas that seemingly have

0:31:570:31:58

nothing to do with driving are busy making

0:31:580:32:02

friends with it, unwittingly

0:32:020:32:04

preparing AI for its upcoming automotive challenge.

0:32:040:32:07

Ruby, for example, has memorised more than 100 billion web pages.

0:32:090:32:13

No human can do this. In fact,

0:32:130:32:15

it can find the right web page as

0:32:150:32:16

you're still typing your search term.

0:32:160:32:19

That is completely impossible for the human brain.

0:32:190:32:22

I mean, before machine learning we

0:32:220:32:23

used to program computers line by line.

0:32:230:32:26

But now, we can teach computers and let them learn on their own,

0:32:260:32:30

the same way people learn.

0:32:300:32:31

At the social media giant Facebook,

0:32:330:32:36

artificial intelligence underpins the whole operation.

0:32:360:32:39

Facebook today could not exist without AI. It's as simple as that.

0:32:410:32:46

Over a billion people use Facebook

0:32:460:32:48

every day and they will load their news

0:32:480:32:50

feed a few dozen times every single day.

0:32:500:32:52

And, you know, if you imagine how

0:32:520:32:54

many people it would take if you were to

0:32:540:32:56

line out all the pieces of content that are available to you every

0:32:560:32:59

day, for how to sort out how

0:32:590:33:00

relevant is this story going to be to

0:33:000:33:03

this person, you multiply that by a billion people,

0:33:030:33:05

you see that for a human,

0:33:050:33:07

this would be a task that is absolutely impossible to do.

0:33:070:33:10

Software engineers are trying to achieve artificial intelligence by

0:33:110:33:14

programming computers to process information like the human brain.

0:33:140:33:18

They have come up with artificial neural networks,

0:33:200:33:23

systems that allow what researchers call deep learning.

0:33:230:33:26

My name is Yann Lecun, I run and I

0:33:280:33:30

build a big research group at Facebook,

0:33:300:33:32

discovering new things that nobody has thought about before.

0:33:320:33:35

So when you talk to your phone and it recognises your voice,

0:33:350:33:38

it uses deep learning. When you

0:33:380:33:39

upload your photos on Facebook and Google

0:33:390:33:41

and you can index them and search them,

0:33:410:33:43

they've been recognised by deep learning.

0:33:430:33:46

Deep learning mimics the way we learn ourselves.

0:33:480:33:51

In our brains, this is done by

0:33:530:33:55

strengthening and weakening neural connections.

0:33:550:33:58

Yann has designed an artificial

0:33:580:34:00

system that does a similar thing in the virtual world.

0:34:000:34:03

When you get an image in your eyes, on your retina,

0:34:080:34:10

it goes to the back of the brain and then it's kind of processed by

0:34:100:34:12

multiple layers of neurons.

0:34:120:34:14

And that process takes about 100 milliseconds and it goes through

0:34:140:34:17

only a few layers of neurons, maybe between ten and 20.

0:34:170:34:19

So these neural nets that we're building,

0:34:190:34:21

they're called convolutional networks.

0:34:210:34:24

It's a little bit my invention and

0:34:240:34:25

they're organised in a very similar way.

0:34:250:34:28

The reason that people are excited about Yann's convolutional,

0:34:290:34:32

artificial neural networks is their

0:34:320:34:34

astonishing abilities in object recognition.

0:34:340:34:37

They can be taught, for example,

0:34:390:34:40

to look at the digital data from a photograph

0:34:400:34:43

and let you know whether or not the picture contains a dog.

0:34:430:34:46

This is done by showing the system lots of pictures of dogs,

0:34:490:34:52

while telling the artificial neural network that these are dogs.

0:34:520:34:55

By looking for common features in the pictures,

0:34:570:35:00

the system will develop its own definition of what constitutes dog,

0:35:000:35:04

so that eventually,

0:35:040:35:05

it will be able to spot a dog or the

0:35:050:35:07

fact that a picture doesn't contain

0:35:070:35:09

a dog, all by itself.

0:35:090:35:11

Of course, it's not just dogs. It could be anything - cars, people,

0:35:120:35:17

penguins, oncoming traffic.

0:35:170:35:19

It's actually better at doing this

0:35:210:35:24

than a human engineer would be at designing a system.

0:35:240:35:27

That's the surprising thing, it's

0:35:270:35:28

very humbling for an engineer to think that's the case, but it is.

0:35:280:35:31

And they work really well for image

0:35:310:35:32

recognition, for video interpretation.

0:35:320:35:35

So all the self-driving cars that

0:35:350:35:37

you see around that use a camera input,

0:35:370:35:39

they all use convolutional nets.

0:35:390:35:40

It's these systems' ability to

0:35:400:35:42

differentiate at a better than human level

0:35:420:35:44

which is at the heart of artificial intelligence.

0:35:440:35:46

It's what powers Brett,

0:35:480:35:50

the knot-tying robot, and it's going to be critical to the difference

0:35:500:35:53

between level three driverless cars that almost work

0:35:530:35:56

and level four driverless cars that really do.

0:35:560:35:59

But it's not the only thing.

0:36:010:36:02

The other thing is this.

0:36:030:36:05

Video games are massively demanding on computer power,

0:36:060:36:10

so much so that a whole new way of

0:36:100:36:12

processing is needed to handle all the data they use.

0:36:120:36:15

What evolved was this - the graphics processing unit or GPU.

0:36:170:36:22

And because of their data handling capabilities,

0:36:240:36:27

GPUs have been gleefully adopted by

0:36:270:36:29

the would-be makers of driverless cars.

0:36:290:36:32

My name's Danny Shapiro.

0:36:340:36:36

I'm the Senior Director of Automotive at Nvidia

0:36:360:36:38

and we're building systems

0:36:380:36:40

to enable cars to drive themselves.

0:36:400:36:42

This is the brain of a self-driving car.

0:36:440:36:47

We plug in cameras, radar,

0:36:470:36:49

lidar. All the sensors of the car feed into this device.

0:36:490:36:54

A CPU, which is the central processing unit,

0:36:550:36:57

you've probably heard, has dual core or quad core,

0:36:570:37:02

meaning there's two lanes or four lanes where information flows.

0:37:020:37:07

The GPU can have thousands of cores or lanes.

0:37:070:37:10

Imagine a highway with 1,000 lanes,

0:37:100:37:13

how much traffic could you push through that processor?

0:37:130:37:15

Just like the brains of human drivers,

0:37:200:37:23

the systems that control driverless cars will be

0:37:230:37:27

voracious consumers of data.

0:37:270:37:29

They will be fed with a constant stream of digits from lidar, radar,

0:37:290:37:35

infrared sensors and multiple video cameras,

0:37:350:37:37

all of which will need to be seamlessly interpreted,

0:37:370:37:40

coordinated and fed back in the form of different data to the car's

0:37:400:37:44

driving controls in real time.

0:37:440:37:47

The driverless car has to use all this to sense and interpret the real

0:37:470:37:51

world with 100% accuracy.

0:37:510:37:54

A GPU is able to process and reconstruct, essentially,

0:37:570:38:00

a three-dimensional model of everything going on around the car.

0:38:000:38:03

All that data, then, is analysed.

0:38:030:38:06

It doesn't just sense there's an object,

0:38:060:38:08

but we know exactly what that object is.

0:38:080:38:10

It could be a pedestrian on a cellphone,

0:38:100:38:13

it could be a motorcycle, it could be an ambulance.

0:38:130:38:16

British start-up FIVE AI are training

0:38:160:38:18

their car to recognise things too.

0:38:180:38:20

Formula 1 engineer Teena Gade

0:38:210:38:23

is being shown the world through the eyes of the driverless car.

0:38:230:38:26

So talk me through what we've got on the screen here.

0:38:280:38:31

The computer vision is making the machine see.

0:38:310:38:33

So here we have the live stream from the camera coming in and these are

0:38:330:38:38

representations that have been processed from that.

0:38:380:38:41

So in the first one, what we're seeing here is in real time,

0:38:410:38:44

the actual detection of buses, cars,

0:38:440:38:46

pedestrians that have been inferred by our algorithm.

0:38:460:38:50

The next one is what we'd call a segmentation,

0:38:500:38:53

which is a breaking-up of the image into something like here's where the road

0:38:530:38:56

is, here's a wall, here's a building,

0:38:560:39:00

so that the car has a very coarse awareness of what's around it.

0:39:000:39:04

And the same sorts of machine learning techniques,

0:39:040:39:06

neural networks is what will also

0:39:060:39:09

help you solve looking at intentions of

0:39:090:39:11

other road users. You can imagine, just as a human goes out,

0:39:110:39:14

they learn how other road users use by observation.

0:39:140:39:17

So you'd feed it millions and

0:39:170:39:19

millions of days of video in all different situations.

0:39:190:39:23

And the machine itself would learn how to understand their

0:39:230:39:27

movements, maybe picking up cues that to you and I,

0:39:270:39:30

we would never have even thought of,

0:39:300:39:31

because it would have so much data at its disposal.

0:39:310:39:36

But there is a fascinating short cut to this,

0:39:360:39:39

one that Teena exploits in her work.

0:39:390:39:41

You might want to turn round.

0:39:410:39:43

I don't know what's up here.

0:39:430:39:45

And one that FIVE AI are making full use of.

0:39:450:39:48

This, actually, looks quite familiar to me.

0:39:480:39:50

This is a simulated environment,

0:39:500:39:52

and that's how we test our cars on the track.

0:39:520:39:54

Yes. So one of the exciting things about the development in gaming

0:39:540:39:58

engines and simulations of reality is they're getting so good...

0:39:580:40:01

I mean, if you look at, say, Grand Theft Auto and games like this,

0:40:010:40:05

it's now become possible to

0:40:050:40:07

actually do testing in these virtual worlds

0:40:070:40:10

which is almost as good as testing in reality.

0:40:100:40:12

And you could get algorithms up to a

0:40:120:40:14

sufficient case on very rare test cases

0:40:140:40:17

that you just wouldn't have access to.

0:40:170:40:19

You know, creating accidents, things like this,

0:40:190:40:21

which you wouldn't want to do in the real world, and actually, you know,

0:40:210:40:25

it takes a lot of money to run a car out in the real world for, like,

0:40:250:40:28

90 million miles.

0:40:280:40:30

If you have enough computers, you can do it in the virtual world

0:40:300:40:33

in, you know, hours.

0:40:330:40:35

It's the modern blurring of the real

0:40:360:40:38

with the digitally virtual that means

0:40:380:40:41

that now, right now, in the early 21st century,

0:40:410:40:45

it might be that the driverless car's time has come.

0:40:450:40:48

Most of the ideas about AI have been around for quite a while.

0:40:510:40:54

What we now have is the technology.

0:40:540:40:56

So we have the computational power

0:40:560:40:57

to run them as well as the ability to

0:40:570:40:59

gather the data in order to learn

0:40:590:41:01

what we need to learn and to train these systems.

0:41:010:41:03

So we're at this point now where driverless cars and driving

0:41:030:41:06

simulators can do both of these things,

0:41:060:41:08

because the technology enables us to at the moment.

0:41:080:41:11

And that's certainly what the car industry believes.

0:41:110:41:14

Welcome to the world of level four.

0:41:140:41:16

Level four.

0:41:190:41:21

According to the plan,

0:41:210:41:23

a level four car will perform all

0:41:230:41:25

driving functions and monitor roadway

0:41:250:41:28

conditions for an entire trip within

0:41:280:41:30

the operational designed domain of the vehicle.

0:41:300:41:33

So this is it.

0:41:350:41:37

This is, to all intents and purposes, the future...

0:41:370:41:40

..driverless cars that will ferry you around with no need for you to worry

0:41:410:41:45

yourself with troublesome bits of kit like steering wheels, or brakes,

0:41:450:41:49

or gears, or anything, really.

0:41:490:41:53

So just imagine, you go to your smartphone, you say,

0:41:530:41:56

"I need a ride to the pizza place," and you punch in the app,

0:41:560:42:00

and the car pops up in front of your house, empty.

0:42:000:42:03

You hop inside, it drops you at the restaurant, you have dinner,

0:42:030:42:06

you drink a lot, you are reasonably drunk now, OK?

0:42:060:42:10

And you will go home and do the same thing again,

0:42:100:42:12

and the car safely brings you home.

0:42:120:42:14

That's going to happen in the next five years.

0:42:140:42:16

So the future's bright for pizza-eating alcoholics,

0:42:160:42:19

so long as they live in the

0:42:190:42:20

operational designed domain of their vehicle.

0:42:200:42:22

But within these zones, level four does offer complete autonomy,

0:42:240:42:29

with no driving controls available to passengers at all.

0:42:290:42:32

However unlikely this vision of the future may seem,

0:42:340:42:37

this is exactly what some

0:42:370:42:39

manufacturers are promising us is just around the corner.

0:42:390:42:43

We're announcing Ford's intent to have a high-volume, SAE, level four,

0:42:430:42:49

fully autonomous vehicle in

0:42:490:42:52

commercial operation in 2021 in a ride-hailing

0:42:520:42:57

or ride-sharing service.

0:42:570:42:59

It's a bold claim,

0:43:000:43:02

and one that will test the navigational systems to their limit.

0:43:020:43:06

But even if these machines come good,

0:43:060:43:08

there's another subtle aspect to

0:43:080:43:10

driving that is sometimes overlooked.

0:43:100:43:13

Probably because it's such a human issue.

0:43:130:43:15

This small autonomous machine is called Jack Rabot.

0:43:220:43:26

By going about his smartly dressed business,

0:43:270:43:30

he's finding out how people behave around each other.

0:43:300:43:32

Jack Rabot learns from the behaviour of other people how to move around.

0:43:350:43:40

The more he looks at people, the better he will be in his navigation.

0:43:400:43:44

It turns out that we humans, when we navigate in crowded scene,

0:43:450:43:49

we read each other's behaviour, body language to avoid each other.

0:43:490:43:53

We respect personal space, yield right of way, and the culture,

0:43:530:43:58

the way people decide to move around is different,

0:43:580:44:02

because we all have different social behaviour,

0:44:020:44:06

and this behaviour can only be learned from the observation,

0:44:060:44:09

from the data. We cannot define them.

0:44:090:44:12

We cannot

0:44:120:44:14

write the rules that should be

0:44:140:44:15

applied in the UK and in Japan on the same time.

0:44:150:44:19

That's because, however adept Jack

0:44:190:44:21

might be at schmoozing with Californian students,

0:44:210:44:24

abandon him in, say, Dagenham, and he might struggle.

0:44:240:44:28

It's the same with driving.

0:44:280:44:30

The reason is that

0:44:300:44:33

the way Americans drive is not the

0:44:330:44:35

same as French and British people drive.

0:44:350:44:39

The only limiting factor of a level

0:44:420:44:44

four car is that it will only work in predefined areas.

0:44:440:44:48

A level five vehicle will be able to work anywhere,

0:44:490:44:53

whatever the driving conditions and

0:44:530:44:55

whatever the cultural driving conventions.

0:44:550:44:58

It might be, then,

0:44:580:45:00

that Jack is helping to deliver the ultimate driverless nirvana.

0:45:000:45:04

Level five.

0:45:060:45:08

We have now arrived at full autonomy.

0:45:080:45:11

Now, according to the masterplan,

0:45:110:45:13

the car will have capabilities at least equal to a human driver in

0:45:130:45:17

every possible driving scenario.

0:45:170:45:19

What this actually means is what it says.

0:45:220:45:25

Pop in a postcode or, if you're more rugged,

0:45:250:45:27

GPS coordinates, and off you go, anywhere,

0:45:270:45:31

and do anything you like on the way.

0:45:310:45:33

At FIVE AI,

0:45:350:45:37

Teena is hoping to witness the reasonably priced car's first steps

0:45:370:45:41

on the path to this full level five autonomy.

0:45:410:45:44

OK, so this is basically the first day in the real world.

0:45:450:45:47

It is the first day in the real world, indeed.

0:45:470:45:49

Yeah. So we're at a test track here,

0:45:490:45:51

so we're not going to sort of hopefully damage anything.

0:45:510:45:54

Excellent. Shall we go, then?

0:45:540:45:56

Let's go.

0:45:560:45:58

So what I'm going to do here... We're going to just hit return.

0:45:580:46:01

OK.

0:46:010:46:02

We're now off. I've got overall

0:46:030:46:05

control of the car with this dead man's

0:46:050:46:07

handle here, so if there's anything that goes disastrously wrong,

0:46:070:46:10

we can always stop. But as you can see, we're not going very fast,

0:46:100:46:14

actually, we're going about four or five miles an hour.

0:46:140:46:16

So it's... There's not a huge danger

0:46:160:46:18

of anything particularly going wrong.

0:46:180:46:20

Within one or two days, we'll be

0:46:220:46:23

going round much more complex tracks,

0:46:230:46:26

and within a few weeks,

0:46:260:46:27

we'll be able to deal with simple kind of junctions and certainly able

0:46:270:46:32

to deal with obstacles.

0:46:320:46:33

Actually turned out to be OK. Looked like it was avoiding those cones.

0:46:340:46:38

It did look like it was avoiding those cones.

0:46:380:46:40

Although we weren't fully sure.

0:46:400:46:42

We weren't fully sure.

0:46:420:46:43

It's early days,

0:46:450:46:46

but this particular electric car has got a fair way to go if it's to

0:46:460:46:50

achieve level five autonomy.

0:46:500:46:51

But if it does, it could revolutionise our driving world.

0:46:550:46:59

I'm excited by the prospect of driverless cars.

0:47:020:47:04

You'll just sit in the cab and it will take you where you want to go.

0:47:040:47:07

And to me, that's good. I can do other things.

0:47:070:47:10

Sure, it'd be a big deal to be able to get into your car and curl up in

0:47:100:47:12

the back seat and have a nap while you go from A to B, but

0:47:120:47:16

if you think about how cities and whole countries are built around the

0:47:160:47:19

road network,

0:47:190:47:21

the changes that could happen to that are enormous.

0:47:210:47:24

And that's really the challenge -

0:47:240:47:27

what does the infrastructure need to look like for these vehicles?

0:47:270:47:30

One major challenge is that, for the past century or so,

0:47:360:47:39

we've been building our world around a totally different kind of car.

0:47:390:47:43

And before you even think about autonomy,

0:47:430:47:46

you need to gear up for electric.

0:47:460:47:48

My name's Gareth Dunsmore.

0:47:500:47:52

I run electric vehicles for Nissan in Europe.

0:47:520:47:54

I perhaps enjoy driving

0:47:540:47:56

a Leaf more than any vehicle I've driven,

0:47:560:47:58

an electric vehicle more than any other vehicle I've driven.

0:47:580:48:01

-Really?

-Yeah.

-You just come here, you have to say that, don't you?

0:48:010:48:03

No. It's different, it's a different driving experience.

0:48:030:48:06

If you've not done it you can't explain it, but it's...

0:48:060:48:09

It's instant acceleration,

0:48:090:48:10

and that... You don't just... You don't get that even in a GTR.

0:48:100:48:14

Apart from the obvious thrill of

0:48:140:48:15

driving Nissan's entry-level electric

0:48:150:48:18

option, Gareth has another reason to be evangelical

0:48:180:48:21

about alternative power.

0:48:210:48:23

Vehicles are far easier to automate if they are also electric.

0:48:230:48:27

On top of that, though, I think there's a broader point.

0:48:270:48:30

Looking at cities and looking at what we're trying to achieve,

0:48:300:48:33

it's about moving to zero emissions and zero fatalities.

0:48:330:48:37

So combining the two technologies together makes

0:48:370:48:40

more sense from a customer perspective,

0:48:400:48:43

to be able to bring forward both at the same time.

0:48:430:48:45

In planning for this electric utopia,

0:48:470:48:49

Gareth teamed up with British architects Foster + Partners.

0:48:490:48:53

Together, they came up with a

0:48:530:48:55

version of the future that imagines how

0:48:550:48:57

electric autonomous cars might do more than just drive themselves.

0:48:570:49:01

I think the challenge

0:49:030:49:05

is integrating it into, you know, existing city fabric.

0:49:050:49:08

And that's where the conversations with Nissan about taking one part of

0:49:080:49:11

it and the charging strategy and autonomously doing that...

0:49:110:49:14

They're all going to have their own nuances which need to be solved.

0:49:140:49:17

I think it's shifted our perspective slightly.

0:49:170:49:19

We kind of, you know...probably would have tackled the problem in a

0:49:190:49:23

very, very different way. But you realise

0:49:230:49:25

the issues facing urban development around the world,

0:49:250:49:27

you need a very integrated approach to transport, in particular,

0:49:270:49:31

how you move from public to private transport,

0:49:310:49:33

how you cover big distances, what's the economics of all of that,

0:49:330:49:37

how do you make it work?

0:49:370:49:39

With Nissan's wireless charging and universal connectivity,

0:49:390:49:43

our vehicles could autonomously charge themselves

0:49:430:49:47

and then re-park so another vehicle

0:49:470:49:49

on the street could use the same bay,

0:49:490:49:52

all while you sleep.

0:49:520:49:54

I think the thing that we're quite excited about is

0:49:540:49:57

there's less reliance on having your car always parked outside your

0:49:570:50:00

house. Once we release them, your car can come from round the corner.

0:50:000:50:04

We can start opening up some of these sort of residential areas,

0:50:040:50:07

and they'll feel much, you know...

0:50:070:50:09

It almost goes back to, you know, a hundred years ago, streets

0:50:090:50:12

become actually a really nice place to be outside,

0:50:120:50:14

not in your little sort of hidden gardens.

0:50:140:50:17

And we become more sociable again and

0:50:170:50:18

use these spaces that are more shared.

0:50:180:50:20

Self-driving cars are going to have a huge impact.

0:50:220:50:25

If all cars are self-driving, we can get rid of the streetlights,

0:50:250:50:28

cos they won't need them, they'll be able to navigate through complex

0:50:280:50:31

intersections with no collisions.

0:50:310:50:33

Right now, we have in the United States 100 million cars.

0:50:340:50:37

They are parked 97% of the time, only driven 3%.

0:50:370:50:41

So in the future, we're going to have less traffic,

0:50:410:50:44

and once we have robotic, self-driving car taxi services,

0:50:440:50:48

we don't need parking spaces any more.

0:50:480:50:50

It means that the cities will look nicer.

0:50:500:50:52

If cars are not crashing, we're not going to need as many doctors,

0:50:520:50:56

and we can totally reimagine the car.

0:50:560:50:59

We don't need the heavy steel, rigid bodies of these cars to protect the

0:50:590:51:03

inhabitants. They could be made of more environmental friendly,

0:51:030:51:06

more lightweight materials.

0:51:060:51:08

So great societal benefits with self-driving cars.

0:51:080:51:12

Well, that's settled, then.

0:51:140:51:15

Driverless cars will usher in a

0:51:150:51:17

world where road traffic accidents will be

0:51:170:51:20

a thing of the past, where the lion will lie down with the lamb,

0:51:200:51:23

swords will be beaten into ploughshares,

0:51:230:51:26

and people will be nicer, kinder,

0:51:260:51:28

happier and richer.

0:51:280:51:30

Except, of course, they won't.

0:51:310:51:33

I think a wholly autonomous,

0:51:340:51:36

driverless future could have more social impacts than we imagine.

0:51:360:51:40

I mean, first off, you've got to

0:51:400:51:41

think about all the people who drive to make a living.

0:51:410:51:44

And so you've got to think about how automation will affect

0:51:440:51:46

their livelihood.

0:51:460:51:49

There are about half a million taxi drivers, delivery drivers

0:51:490:51:52

and bus drivers in the UK,

0:51:520:51:54

who might well need to look for alternative employment.

0:51:540:51:57

But as if that wasn't bad enough,

0:51:580:52:00

driverless cars might even make our

0:52:000:52:02

cities' congestion problems even worse.

0:52:020:52:04

I can see very quickly a time where people won't actually stop their car

0:52:050:52:09

driving when they want to go to the shops.

0:52:090:52:11

They'll go to the shops, they'll get out,

0:52:110:52:13

and then they'll tell their car to drive around the block and wait till

0:52:130:52:15

they're ready.

0:52:150:52:17

And the problems don't stop there.

0:52:170:52:20

There are practical and ethical issues yet to be resolved.

0:52:200:52:23

It'll be difficult for autonomous cars to coexist with people driving

0:52:240:52:27

cars, which will obviously be the case for many years to come.

0:52:270:52:31

The cars with drivers are

0:52:310:52:32

potentially going to take advantage of the

0:52:320:52:35

driverless cars, which they know have to stop and give way when their

0:52:350:52:38

sensors detect that something's wrong.

0:52:380:52:40

What happens when it eventually encounters a no-win scenario,

0:52:430:52:46

when it actually has to have an accident, where it cannot avoid it?

0:52:460:52:50

If you get to the point where the car has to choose between

0:52:500:52:56

an old man in a vehicle or a young lady and an infant in the other car,

0:52:560:53:01

and it knows it's going to hit one of them,

0:53:010:53:03

these cars are going to have to be taught to make these moral,

0:53:030:53:06

ethical judgments that you make in a snap decision.

0:53:060:53:10

Some car manufacturers have already said that they're going to teach the

0:53:100:53:13

car to favour the driver above all else.

0:53:130:53:17

Now, what if you don't want that?

0:53:170:53:19

You might actually set those parameters in the vehicle yourself,

0:53:190:53:22

the level of morality for the car.

0:53:220:53:24

Where some driver-assist technologies have been introduced,

0:53:270:53:30

there have already been problems.

0:53:300:53:33

A Tesla crashed, killing its driver,

0:53:330:53:35

and Uber suspended their fleet of

0:53:350:53:37

semi-autonomous test cars following a non-fatal collision.

0:53:370:53:42

But despite this, the technology moves forward.

0:53:420:53:45

It's three weeks since Teena was at FIVE AI, and today,

0:53:530:53:56

she's going back to meet them for the last time,

0:53:560:53:58

to see how their car's getting on

0:53:580:54:00

with its assault on level five autonomy.

0:54:000:54:02

What we're now going to do is just engage our robot.

0:54:060:54:09

Press that button there. Robot is now in control.

0:54:090:54:12

And off we go.

0:54:120:54:14

OK, so the car's now setting off.

0:54:140:54:17

See it's quickly got up to...

0:54:170:54:19

Doing about 12 miles an hour on this access road here.

0:54:190:54:21

So right now, we're driving entirely autonomous.

0:54:210:54:23

We're driving entirely autonomously.

0:54:230:54:25

Actually, this robot's getting a bit rattly.

0:54:250:54:28

-It is.

-A few of the things we've

0:54:280:54:30

really improved over the last few weeks

0:54:300:54:32

is we've managed to sort out a lot of the visual odometry,

0:54:320:54:35

so the car can work out really

0:54:350:54:36

accurately where it is in three-dimensional

0:54:360:54:39

space. Just from the cameras and the pixels on the camera and their

0:54:390:54:43

-motion...

-Yeah.

-..it can actually work out...

0:54:430:54:45

It can in fact go round really quite long tracks,

0:54:450:54:48

and within a few centimetres can detect exactly where it is in

0:54:480:54:51

three-dimensional space without the use of a map.

0:54:510:54:54

-Yeah.

-So that's pretty cool.

0:54:540:54:56

Although, not quite cool enough.

0:54:560:54:59

An old-fashioned level zero

0:54:590:55:01

interception is required to save the car from itself.

0:55:010:55:04

Whoa, it's too close.

0:55:040:55:06

Let's just make sure this thing is not going to hit that barrier there.

0:55:060:55:10

There's some latencies in the system that we're now going to iron out.

0:55:100:55:13

And when you say latency, what's actually happening

0:55:130:55:15

is the car's not responding as fast as you'd like.

0:55:150:55:17

That's right. Yeah, it takes...

0:55:170:55:19

It takes a few hundreds of

0:55:190:55:21

milliseconds to go all the way through from

0:55:210:55:23

seeing data in the cameras to

0:55:230:55:25

actually a decision that actually controls the wheel.

0:55:250:55:29

And that gap between the two

0:55:290:55:31

means that by the time we actually apply control,

0:55:310:55:34

it's to a situation that was true maybe half a second ago.

0:55:340:55:37

-Right.

-And, you know,

0:55:370:55:39

we need to take all that into

0:55:390:55:40

account in the way we design the car.

0:55:400:55:42

Also the way we speed up some of the algorithms, as well.

0:55:420:55:44

-Yeah.

-Stop it there at that point, actually.

0:55:440:55:47

We're about to hit the roundabout. So just go round this roundabout.

0:55:470:55:49

Basically, it is still in a slightly drunken manner, but...

0:55:490:55:53

Yeah, it is... It is suffering from a serious case of PIO.

0:55:530:55:56

Pilot-induced oscillation.

0:55:560:55:58

Responding to the signal too late.

0:55:580:56:00

Very well documented in lots of aeronautics texts.

0:56:000:56:02

Well, there you go. You know,

0:56:020:56:04

this is why we should have cross-disciplinary...

0:56:040:56:06

Yeah, you need some dynamicists on board.

0:56:060:56:08

We do need some dynamicists, I tell you. Yeah.

0:56:080:56:12

Despite the obvious challenges, Stan remains optimistic.

0:56:120:56:15

In 2019, we're going to be driving autonomously in urban scenes,

0:56:180:56:22

and I reckon there's going to be commercial, you know,

0:56:220:56:25

within one-and-a-half to two years after that.

0:56:250:56:27

FIVE AI have chosen the hardest problem to solve,

0:56:280:56:32

that of complete autonomy,

0:56:320:56:34

so their journey in reaching that goal could well be a long one.

0:56:340:56:37

The prospect of replacing our cars

0:56:440:56:46

with driverless ones tends to split opinion.

0:56:460:56:49

You know, we've done sort of surveys

0:56:510:56:52

and focus groups and things like that

0:56:520:56:55

with people, and it's honestly 50-50.

0:56:550:56:58

Some people hate the idea of

0:56:580:57:00

relinquishing the pleasure of driving.

0:57:000:57:02

I love driving. I'm an automotive journalist,

0:57:020:57:05

so I have to love driving or I wouldn't do my job.

0:57:050:57:07

Others relish the prospect of

0:57:070:57:09

freeing up wasted time behind the wheel.

0:57:090:57:12

If it can make my commute into work that much easier,

0:57:120:57:15

I won't have to stress about it in the morning, that'd be grand.

0:57:150:57:18

If I could just sit back and read a book, listen to some music,

0:57:180:57:21

catch up on some sleep, that would be great.

0:57:210:57:24

For some, it's the ultimate freedom.

0:57:240:57:26

For others, it's an opportunity for less freedom, for a longer work day.

0:57:260:57:30

All those times when you were travelling,

0:57:320:57:34

and it was a break from the phone calls and it was a break from the

0:57:340:57:38

e-mail, I think those times would be freed up for us to actually carry on

0:57:380:57:43

working and the daily grind.

0:57:430:57:46

But despite the reservations, the die seems to have been cast.

0:57:470:57:52

The dawn of the driverless car is here.

0:57:520:57:55

The challenge for the technologists is to make sure that the transition

0:57:550:57:59

into reality is as beguilingly smooth as the PR that surrounds it.

0:57:590:58:04

I can't foresee the future, but I can build it.

0:58:040:58:08

The reason that I believe it's good to be a technology optimist is

0:58:080:58:11

throughout the entire history of the human race,

0:58:110:58:13

technology has empowered us.

0:58:130:58:15

From the very early days, the Bronze Age, the Stone Age,

0:58:150:58:19

to the day of the smartphones and modern medicine, it has freed us,

0:58:190:58:23

it has levelled the playing field for everybody,

0:58:230:58:25

and has empowered us as a human race.

0:58:250:58:27

Why stop that?

0:58:270:58:28

Download Subtitles

SRT

ASS