Media release

From: Springer Nature

Technology: AI can outrace champion Gran Turismo drivers (N&V)

An artificial intelligence (AI) agent that can outcompete world champion-level human players in the head-to-head car racing game Gran Turismo is reported in a Nature paper. The agent demonstrated exceptional driving speeds, car control and tactics whilst maintaining racing etiquette. The findings could have useful applications for autonomous navigation and fundamental AI research.

Many potential applications of AI require making real-time decisions in physical systems whilst interacting with humans. Car racing represents an example of these conditions; drivers must execute complex tactical manoeuvres to pass or block opponents whilst operating their vehicles at their traction limits. Racing simulations, such as the PlayStation game Gran Turismo, reproduce the control challenges of real race cars and provide a challenging application for machine learning.

Peter Wurman and colleagues taught an AI agent, named GT Sophy, to play Gran Turismo using deep reinforcement learning. The agent was trained to master the art of accelerating and braking efficiently over race courses, and learned how to find alternative paths in different conditions or when blocked by opponents. One of the most challenging aspects for the training of a successful AI is to ensure it avoids incurring penalties associated with breaching race etiquette, a set of loose rules adjudicated by external human judges. GT Sophy was able to win head-to-head competitions against four of the world’s best e-sport drivers using three car and track combinations, each posing different racing challenges, including vehicles that can reach speeds in excess of 300 km per hour.

These results are another step in the ongoing progression of competitive tasks that computers can beat the very best humans at, such as chess and poker. The findings could make racing games more enjoyable and provide realistic high-level competition for training professional drivers and discovering new racing techniques, the authors suggest. The approach may also find applications in real-world systems, such as robotics, aerial drones and self-driving vehicles.