One of the best things about computers is that they can learn as much from a simulation as they can from so-called ‘real world’ experiences. That means, given the right simulator, we can learn AI to drive cars without ever putting a single human being in danger.
Almost all AI companies train their driverless vehicle algorithms using simulations. Until now, the simulators themselves were not that interesting. They are mostly just physics engines designed to be interpreted by a neural network. But Sony just unveiled the most popular autonomous driving simulator ever: Gran Turismo Sport.
If you are not a player: this is not advanced software designed to train AIs, it’s a game. And not just any game but the latest in one of history’s most beloved racing simulation series
Researchers from the University of Zurich and Sony AI Zurich recently published a pre-printed paper showing the development of an autonomous agent designed to beat the best human players in the game.
Among racing games, the Gran Turismo Sport (GTS) is known as a highly realistic driving simulation that models phenomena such as the influence of tire temperature and the car’s current fuel level on traction. Therefore, the optimal track (ie the track that leads to the fastest lap time) for a car in the GTS, like real world racing, depends not only on the geometry and characteristics of the track, but also on different (a priori unknown) car physical characteristics and conditions. Due to its similarity to real driving and the relatively low cost of training in the GTS compared to training with actual race cars, the GTS is also used to throw drivers to racing teams.
In other words: It’s a legitimate simulation used by real – world racing teams to help determine the capabilities of real, expert level rules. That’s pretty high praise for a video game.
The researchers had a fairly high order to fill. While AI systems regularly outperform humans in games like Chess and Go, ordinary computer-controlled racers tend to be mean to human experts.
The researchers write:
As far as we know, the built-in non-player characters (NPCs) are included in modern racing games is unable to compete with human expert players in reasonable comparisons. For example, the current one built-in NPC in Gran Turismo Sport (GTS) loses a total of 11 seconds compared to the fastest human driver and is slower than 83% of all people in one of our reference settings.
Other racing games apparently close the gap for human experts by giving the NPC an unfair advantage, for example, by increasing the engine power of the NPC’s car; however, this leads to frustration among human players who feel cheated.
Instead of cheating or adapting the rules, the team turned to a facet of AI called deep reinforcement learning. This involved training AI to recognize the way forward and respond in a more humane way.
According to an article by Tech Xplore author Ingrid Fadella, Yunlung Song, a co-author of the team’s research paper, said:
Unlike classical condition estimation, path planning, and optimal control methods, our approach is not dependent on human intervention, human expert data, or explicit path planning. We found that it could generate tracks that are qualitatively similar to those selected by the best human players, while surpassing the best known human lap times in all three of our reference settings, including two different cars on two different tracks .
As far as we know, this is the first time that an autonomous car AI beats human experts in Gran Turismo Sport. And while there is currently no artificial intelligence system capable of level five autonomy (capable of driving a vehicle without external aids or human assistance), if you absolutely must drive a vehicle controlled by an AI: may as well choose the one trained in a video game to push the physical limits of speed and control.
You can read the whole paper here.
So you are interested in AI? Then Attend our online event, TNW2020, where you will hear how artificial intelligence is transforming industries and businesses.
Published September 15, 2020-22: 03 UTC