Self-driving cars are the future and seemingly every tech company on the planet is hungry to get in on the action. But navigating the world's complex urban and suburban environments will be a challenge for these machines and innovation will come in fits and starts. Less complicated and more likely to be a starting point for the driverless revolution is the highway system and the trucks that ship things from point A to point B.
Mars Auto, a new South Korea-based startup, is aiming to pioneer such technology. To that end, its developers created a training program for its truck driving artificial intelligence that uses the popular PC game Euro Truck Simulator 2. Mars Auto is using a video game to teach trucks how to drive by themselves.
"Before we tested our program on the road, we wanted to test it in games," Gyuri Im, one of Mars Auto's founders, told me via Skype.
Gyuri and his team aren't the first to do something like this. As he pointed out in a Medium post about the project, a game called The Open Racing Simulator (TORCS), is a popular choice for companies looking to use a game to develop self-driving software. But Gyuri and his team thought they could do better than TORCS.
"It's not really close to the real world," he said. "It's a racing game. We wanted something that's more realistic in terms of physics and environment."
Mars Auto picked Euro Truck Simulator 2 because the startup is interested in transportation and logistics. The game was the closest thing to a virtual reality simulation of a complex highway system it could find.
When training an AI, Gyuri explained, the closer you can make the simulation to the real world the better. Games are great, in principle, for training an AI because "it can generate a lot of training data instantly in a virtual environment," Gyuri said. Euro Truck Simulator 2 in particular is a good fit for this purpose because it was designed to simulate the real world, as opposed to a game like Grand Theft Auto V, which at times bends realistic physics to make the game more fun.
Using a video game to train a deep learning algorithm is just one part of what will be a long and complicated process. The game works for Mars Auto's purposes now, but there are limitations. "If we want to develop more sophisticated algorithms, we can't do it with Euro Truck Simulator," Gyuri said. "Self-driving technology uses other sensors beyond the camera [like] radar and sonar. You can't simulate that input using Euro Truck Simulator 2."
For its next step, Mars Auto is looking at building its own game and including simulations of radar and sonar to train up the next evolution of its self-driving truck software.