You are driving down the street and see a car coming towards you, swerving in and out of your lane. It is being followed by four or five equally erratic police cars. What do you do? Do you slow down? Or pull over to the side and stop? It is a split-second decision and, for most people, a terrifyingly new experience.
Such a scenario is a problem for humans, but it is even more of a problem for self-driving cars. A self-driving algorithm compares the information it receives from imaging, radar, or LIDAR with previous input, and when confronted with a new scenario, it can only react based on these computations. A human can say “If I pull over, I can avoid the chase. I think I remember seeing that in a movie once.” The algorithm does not have the same ability.
This is why some makers of self-driving cars are turning to video games, which can replicate ‘n’ number of likely and unlikely scenarios in order to enhance machine learning. Over-the-top games like the ‘Grand Theft Auto’ series are being used to quickly increase the knowledge base for self-driving cars as video games simulate the real world. This is an example illustrating that OEMs and tier-1 suppliers are willing to look for innovative, alternative solutions to the challenges posed by new technology.
Video Game Learning and Artificial Intelligence
Google has used games to test artificial intelligence (AI) capabilities for years. Last year, its AI system mastered ‘Space Invaders’, a classic game for the Atari 2600. But the same system struggled with ‘Pac-Man’, for a couple of reasons: one – it was unable to plan more than a few seconds in advance, which meant it could not draw ghosts away and then double back to collect pellets, and two – it could not figure out that some of the pellets made ghosts vulnerable. The fact that ghosts changed color when they could be attacked did not mean anything to the AI system as this was outside its scope of knowledge.
In other words, straightforward shooters like ‘Space Invaders’ could be mastered, but even the most basic thinking games were a challenge. If computers cannot actually think creatively, they can still learn to mimic the effect by experiencing every potential possibility and drawing on this information to create a solution. Especially with modern video games, there is an almost infinite range of possibilities. Take the game, ‘No Man’s Sky’, which is programmed to create new worlds and galaxies based on gameplay, it is estimated that it would take 548 billion years to play through every possible scenario.
The benefits are clear – when given the chance to play through endless simulations, AI machines will learn more about how the world works. AI is essentially a huge data-processing program and in order to be functional, to pass the Turing Test, it needs as much input as possible.
‘Grand Theft Auto’ and the Endless, Chaotic Road
When it comes to autonomous cars, in particular, the biggest challenge is recognizing everything on the road. A computer can be told what humans look like, of course, but can a moving car recognize a hunched-over person in a bulky coat running into the street after a wind-blown $20-dollar bill? Or an overflowing pickup truck that is one jolt away from spilling pipes and tools onto the highway?
It is impossible to program every possibility. This is why self-driving cars currently rely on fleet learning, where, thanks to Over-The-Air (OTA) connectivity and software updating capabilities, one car can ‘learn’ something and every vehicle in the fleet will receive this updated information. This is, however, a slow way to learn, which is where ‘Grand Theft Auto’ comes in. The game essentially consists of absurd car chases through meticulously rendered streets. The player is encouraged to drive like a reckless maniac, with no concern for the law, other drivers, or pedestrians, so actually learning to drive based on input from the game is a terrible idea, but that is not what the self-driving AI is doing.
Instead, it is learning to recognize objects – such as trees, buildings, and erratic cars – which it might encounter in the real world. The simulation can play hundreds of games at once, going through nearly every scenario imaginable. It can create a comprehensive picture of every possible world. This allows for much faster learning. They will be creating simulations enhancing machine learning, even if (perhaps especially if) the simulations are extreme. It is relatively easy to program a car to drive in regular traffic. It takes extreme measures to ensure it is ready to respond to a police chase after a bank heist or any other corner case it may encounter.
That is what OEMs will need to do. They must be willing to remain open-minded in order to find innovation from the least likely sources. OEMs will soon become gatekeepers for the best new ideas, and it is essential that they use their expertise to test and vet all possible self-driving solutions before extending them to their fleets. This is the best way to keep everyone safe, and ironically, as far away from the mayhem-filled streets of ‘Grand Theft Auto’ as possible.
As the auto industry is changed by technological and economic currents, OEMs and Tier-1 manufacturers will need to partner with technological specialists to thrive in the era of the software defined car. Movimento’s expertise is rooted in our background as an automotive company. This has allowed us to create the technological platform that underpins the future of the software driven and self-driven car. Connect with us today to learn more about how we can work together.