23 November 2017

Self-driving cars

GTC 2017

miscellaneous | Self-driving cars are the future. Everyone agrees on that. But when will autonomous cars actually appear on the road? And how do they work? There's a lot of confusion about that, to put it mildly. Autozine found the answers at "GTC 2017", a worldwide conference for developers of self-driving cars.
 

The ultimate self-driving car can steer, brake and accelerate all by itself. A driver is no longer necessary. And that solves a lot of problems. Reading, playing with a mobile phone, watching TV or sleeping are no longer prohibited but rather the most logical things to do while driving.

And there's more good news. A car that's being driven by a computer, is more accurate than a human being. Autonomous cars have perfect vehicle control and can drive closer together, thereby utilising the roads more efficiently. Also, they are never distracted, have a perfect 360 degree view, always obey the rules and never make mistakes. This means the end of traffic victims, traffic jams and daily frustrations.

Things will progress even further when cars slowly turn into "robot taxis", which after every ride go and look for a new passenger. This will drastically reduce the number of cars on the road. Problems with the environment will mostly be solved and parking lots can be transformed into parks or playgrounds.

GTC 2017

So far for the future. During "GTC 2017" (GPU Tech Conference) software developers, economists and carmakers came together to discuss and learn exactly with this future in mind.

GTC 2017 was organised by "nVidia". This company is best known for graphical chips inside computers.  PCs used for 3d gaming require specialised "graphics cards" which can render at mind-blowing speeds.

Gamers demanded ever faster graphics cards from nVidia and in the end the computing power was so great that scientists started using this capacity for other means. Instead of a super computer, they deployed a games computer for the heavy mathematical lifting!

With the introduction of the self-driving cars, the roles are being reversed. Before, the graphics card showed the virtual world in which the gamer controlled a car. Now the image comes from cameras in the real world and the computing powered is used to interpret these images by the computer that drives the car.

Artificial intelligence

"interpret the images" is easier said than done. A computer can easily recognise colours and patterns, but understanding is much more difficult. Recognising a traffic sign by its shape and colour is no problem. Recognising complex objects like cars or pedestrians is much more difficult, especially when they vary in colour and shape.

The solution is "artificial intelligence" or "machine learning". This means the computer learns from many examples and it is told what's good and bad ("don't drive over the pedestrian", "stop for a red light", etc). Instead of working with an infinite list of all possible traffic signs, all existing cars and all people, the computer learns about generic characteristics of a traffic sign, car or pedestrian. This knowledge is stored in a so-called "neural network".

The more examples an artificially intelligent system sees, the more accurate it gets. An important new step is that nVidia developed a bespoke processor for self-learning systems. The new "Tensor RT" is capable of interpreting 300,000 miles of traffic situations in just 5 hours. Now that's a quick student!

This means that even now the autonomous car is safer and more predictable than even the best human driver. Questions like: "who is responsible in case of an accident?" or "if the self-driving car has to sacrifice either a child or an older person to avoid an accident who will it choose" are therefore irrelevant. The fact is that the self-driving car already does better than humans and will save lives right now.

The self-driving cars

The next step is to transfer the neural network containing all knowledge of driving into the car. Up to now, this required a boot full of computers. During GTC 2017 nVidia introduced a computer the size of a licence plate that takes over all of that functionality. Power consumption also reduced significantly, which means an extra power supply is no longer required.

Modern cars already have cameras, radars, electric power steering and more to assist the driver in dangerous situations. Therefore it's merely a change in software to make the step from assisting to taking over driving. With the introduction of electric vehicles this is even easier, because the computer doesn't have to control a clutch or gearbox.

Testing

So, what are we waiting for? Besides many technical lectures about the "how", speakers from the industry and politics gave a clear answer to the "when" question during GTC 2017.

The first problem is politics. There's a fixed list of requirements the parts of conventional cars have to comply with in order to be allowed on public roads. Because it's hard to understand what's going on inside a neural network of an autonomous car, the only alternative is to run many tests. The industry is waiting for governments to compile such a set of rules and tests for self-driving cars in order to be allowed on the roads.

To demonstrate what self-driving cars are capable of today, several companies showed prototypes at GTC 2017. Most appealing is the "roboracer", which proves that computers have perfect vehicle control. Deutsche Post showed off a van that follows the postman autonomously. The mailman only gets into the car when the distance to the next address is too great to walk.

Industry

The second problem is the industry. Most well-known carmakes have no interest in self-driving cars. Autonomous cars will also be shared cars. This means the number of cars sold will drop dramatically. Also, self-driving cars will be electric cars and that means a minimum of service and maintenance. Not only the carmakers, but the car dealers also have no interest in electric cars or self-driving cars that need no maintenance and no repairs.

That's why at GTC 2017 no less than 45 (!) small companies show how things could be. A financial expert emphasised that these new companies should be taken very seriously. As an example he mentioned Nokia and Ericsson. Those companies once stated that a computer company like Apple could never make a decent phone. Suffice to say they "sort of" misjudged that fruity computer company.

Big carmakers learned from this mistake. That's why they are all developing self-driving cars. Under pressure from the new kids on the block they develop as quickly as necessary, but forced by the powers that be as slowly as possible.

Consumers

The third group that causes the delay, are consumers. They have to trust the technology and eventually release the steering wheel. Just like with the smartphone, eventually the advantages will be obvious. What starts slowly, will soon grow exponentially.

For car fans, life will only get better. For them, driving a car will be like horse riding or sailing: not necessary, but certainly fun to do.

Conclusion

After a week of presentations and demonstrations it's clear: after the smartphone, the smart car will change the world. That will happen step by step. Already cars are fitted with systems that assist the driver. On easy routes and on closed tracks assisting will soon be replaced by taking over. Later still, the car will drive all by itself.

The technology is ready for the self-driving car. Now we have to wait for politics, the industry and consumers to accept the autonomous car. The most pessimistic speaker expects the self-driving car to appear on public roads in 2045. The most optimistic expects 2019. As far as Autozine is concerned the self-driving car will be here as soon as possible, because it will solve a lot of problems and will make the world a better place.