‘Game of Drones’ for the driverless car

For all the talk surrounding autonomous vehicles, little has been said about the technology’s roots in entertainment. Sci-fi movies have showcased the potential of self-driving cars for several years but they haven’t provided any practical solutions. Elements of autonomous driving actually came from another source: video games.

“Some of the earliest games out there were simulators of vehicles and they always had a cause and effect,” said Sanford Russell, senior director of autonomous driving ecosystem in North America for gaming and visual computer specialist NVIDIA. “If I hit the vehicle next to me, it has to go do something or it doesn’t feel like a simulator.”

These simulators were created for entertainment purposes but they are now helping automakers develop autonomous vehicles. “It’s not entertainment now, it’s about safety and convenience,” Russell added. “But the problem was very similar.”

Autonomous vehicles must be able to understand the behaviour of other cars, including turn signals and lane adjustments. Russell said this is no different from the way a game simulator behaves. “The only difference is you’re no longer driving the car – the simulator is,” Russell explained. “Which is the same thing – if you do a simulator, there’s 30 other cars on the track. It’s already driving cars. Now the fidelity and instrumentation we have in an autonomous vehicle is orders of magnitude better than in a game simulator.”

Simulation may also help automakers drive millions of miles without increasing the number of autonomous vehicles being tested on real roads. Russell believes that a huge percentage of that could be handed off to simulators. Thus, for every actual mile driven, autonomous cars could get many more miles worth of invaluable data.

Better with time

As if simulation wasn’t enough, new technologies could allow tomorrow’s vehicles to age more gracefully than the cars that are currently on the market. Said Russell: “This iterative approach to constantly improving is key to autonomous driving. Tesla already does over-the-air updates. They’re able to effectively improve the car after you bought it. If you said that to someone in the auto industry 10 years ago, that was a foreign concept. Now there’s a whole bunch of people out there who enjoy that. They get in it and their car is better! That’s a good feeling.”

Man-made software updates are only the beginning. Deep learning, which is slowly making its way into automobiles, could present new ways for cars to evolve over time. “Think of it as a child,” said Russell. “You have a two-year-old, you throw the ball, the ball hits it in the chest. You throw it again, the arm starts moving. After a month, the hand grabs it. That is training a human the exact same way you train an autonomous car – through repetition. You give it a problem until it understands what it has to do.”

Russell speculated that if an autonomous car is designed correctly, it will perform the way consumers want but self-driving features might be limited to highways until the car learns how to navigate additional environments. He said: “Deep learning is really solving very fundamental problems that were considered unsolvable from a human programmer perspective but through training, we’re finding that neural networks are incredibly powerful.”

Simulation nation

Long before video game tech assisted automakers in their quest for autonomy, car designers used simulators to test a wide variety of elements. “Crash simulation is one,” said Russell. “They want to actually see what happens when a vehicle hits an obstruction. How does it crumple? Are the crumple zones working as expected?”

Automakers also save money by building cars virtually before real automobiles are produced. Said Russell: “You think of design and styling. It used to be that they made these huge clay mules. You’d walk up and look at it. Now they build these huge virtual cars and they can show the car driving at night on a watery road in France. They can change the environment the car is in. The car reflects the virtual environment, so you’re looking at the car in a more realistic way than in an engineering lab where they have a clay mule.”

No robotic fears

Tesla co-founder Elon Musk has been very vocal about his fear that robots will take over. A deep learning vehicle might sound like it could fulfil that fear but Russell isn’t worried. “None of the technology has approached what we call ‘consciousness,’” he said. “Nobody even has that on the roadmap. What people are trying to solve are really fundamental issues. Think of transportation – it sounds simple! It’s only when you start working on it that you realise how hard the automakers are working to solve it. It’s not a trivial problem. The objective you could put on a whiteboard: ‘Take me from Point A to Point B.’ How hard can it be? Well, it’s actually fairly hard.”


Leave a comment

Your email address will not be published. Required fields are marked *