Sony’s new AI agent gets supernatural Gran Turismo Sport score


One of the best things about computers is that they can learn as much from a simulation as they can learn from so-called ‘real-world’ experiences. This means that, given the proper simulator, we can teach AI to drive a car without ever endangering humans.

Just about every AI company trains its driverless vehicle algorithms using simulation. Until now, the simulators themselves were not at all interesting. They are mostly just physics engines that have been interpreted by a neural network. But Sony unveiled the most popular autonomous driving simulator yet: Gran Turismo Sport.


In case you’re not a gamer: It’s not advanced software designed to train AI, it’s a game. And the latest but not the latest in any of the most beloved racing simulation series in history

Researchers at the University of Zurich and Sony AI Zurich recently published a pre-print paper showing the development of an autonomous agent who defeated the best human players in the game.

Per team:

Between racing games, the Gran Turismo Sport (GTS) is known as a highly realistic driving simulation, modeling phenomena, such as the impact of tire temperature and the car’s current fuel level on traction. Therefore, similar to real-world racing, the optimal trajectory (i.e., the path leading to the fastest lag time) for a car in the GTS depends not only on the geometry and properties of the track, but also on the different (one-priority unknowns). ) Also depends on the physical characteristics and conditions of the car. Because of its similarity in actual driving and the relatively low cost of training in the GTS, compared to training with real driving cars, the GTS is also used to select drivers for racing teams.

In other words: This is a valid simulation used by real-world race teams to help determine the abilities of real, expert-level drivers. This is too much praise for a video game.

Researchers had a very tall order to fill. While the AI ​​system routinely outshines humans in games such as chess and go, standard computer-controlled racer specialists perform poorly against human players.

Researchers write:

To our knowledge, modern car racing games include built-in non-player characters (NPCs). Human experts are unable to compete with the players in fair comparison. For example, currently The manufactured 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 humans in one of our reference settings.

Other racing games clearly close the gap for human experts by providing an unfair advantage to NPCs, for example by increasing the engine power of an NPC’s car; However, it creates frustration among human players who feel cheated.

Rather than cheat or bend the rules, the team turned to an aspect of AI called intensive reinforcement learning. In this, AI was trained to recognize the road ahead and react in a more human-like fashion.

According to an article by Tech Xplore author Ingrid Fadella, Yunlung Song, co-author on the team’s research paper, said:

Different from classical state estimation, trajectory planning, and optimal control methods, our approach does not depend on human intervention, human expert data, or explicit path planning. We found that it can produce trajectories that are qualitatively similar to those selected by the best human players, while improving the best known human lap times in all three of our reference contexts, with two on two different tracks Including different cars.

To the best of our knowledge this is the first time an autonomous car AI has beaten human experts In Gran Turismo Sport. And while there is currently no artificial intelligence system capable of level five autonomy (one capable of driving with external assistance or human-assisted vehicles), if you ride in a vehicle completely controlled by an AI Should: So a trained person can also be selected. A video game about pushing the physical limits of speed and control.

You can read the entire paper here.

So are you interested in AI? Again Join our online program, TNW2020, Where you will hear how artificial intelligence is transforming industries and businesses.

Published September 15, 2020 – 22:03 UTC