How GT Sophy Mastered Tire Physics and Slipstreaming

Gran Turismo Sophy (GT Sophy) is a superhuman artificial intelligence agent developed by Sony AI, Polyphony Digital, and Sony Interactive Entertainment. This article explores how GT Sophy used deep reinforcement learning to master two of the most complex aspects of high-speed racing: nonlinear tire friction limits and aerodynamic slipstreaming. By utilizing a specialized off-policy reinforcement learning algorithm, carefully structured reward systems, and multi-agent training scenarios, the AI learned to control cars at the absolute edge of physical grip and execute complex overtaking maneuvers against the world's best human drivers.

The Deep Reinforcement Learning Framework

To conquer the complex physics of the Gran Turismo simulator, researchers created an asynchronous, off-policy training algorithm called Quantile Regression Soft Actor-Critic (QR-SAC). Traditional reinforcement learning evaluates the expected average reward of an action. In contrast, QR-SAC estimates the entire probability distribution of future rewards. This distributional approach allowed GT Sophy to accurately assess the high risks of racing at the limit, such as spinning out or colliding with another car, and balance them against the reward of faster lap times.

Unlike human drivers who see the game screen, GT Sophy perceives the race through raw numerical telemetry. Its neural networks receive a continuous stream of data, including the car’s velocity, acceleration, tire slip angles, and radar-like measurements of the upcoming track boundaries and nearby opponents. It outputs continuous control commands for steering, throttle, and braking.

Mastering Nonlinear Tire Friction Limits

A race car's tires only have a finite amount of grip. The relationship between the forces acting on a tire and the grip it produces is highly nonlinear. As a car approaches its physical limit, the tires begin to slip, which initially increases grip up to a peak threshold. If the driver pushes just slightly past that peak, the tire loses traction entirely, resulting in a sudden spin or skid.

GT Sophy was not programmed with the mathematical equations of tire physics. Instead, it mastered this nonlinear friction curve through trial and error across millions of simulated laps. The key to this mastery was the AI’s reward function, which primarily rewarded progress along the center line of the track while penalizing off-track excursions and tire skidding.

Through this reward structure, GT Sophy learned to balance longitudinal forces (braking and acceleration) and lateral forces (cornering), a concept known in racing as the "friction circle." By receiving immediate feedback on the car's slip angles and velocity vector, the AI learned how to trail-brake—gradually releasing brake pressure while turning into a corner—to shift the car's weight and maximize the contact patch of the front tires. It learned to hold the tires at the exact peak of their nonlinear grip threshold, adjusting throttle and steering inputs in milliseconds to prevent catastrophic traction loss.

Exploiting Aerodynamic Slipstreaming

While mastering tire limits is enough for solo time trials, pack racing requires an understanding of aerodynamics, specifically slipstreaming (or drafting). When a car travels at high speed, it punches a hole in the air, creating a low-pressure wake behind it. A trailing car that enters this wake experiences reduced aerodynamic drag, allowing it to achieve higher speeds and overtake the leading car.

However, slipstreaming introduces a severe physical trade-off known as "dirty air." While the trailing car gains straight-line speed, the turbulent, low-pressure air reduces aerodynamic downforce over its wings and body. This loss of downforce directly reduces the vertical load on the tires, shrinking their available friction limits. If a driver enters a corner in the slipstream without adjusting their braking point, the car will understeer and crash.

GT Sophy learned to exploit slipstreaming and manage dirty air through mixed-scenario training. Sony AI placed the agent in multi-car environments against a variety of opponents, ranging from slow, predictable AI to aggressive, fast AI. The input data included the relative positions, velocities, and accelerations of nearby cars within a specific radar range.

The reward function in these scenarios was expanded to include passing rewards for overtaking opponents and heavy penalties for collisions. Through QR-SAC, GT Sophy associated the spatial position directly behind an opponent with an increase in its own acceleration. It naturally deduced the aerodynamic benefit of the draft.

Simultaneously, the AI learned the danger of dirty air. When attempting to brake late while tucked behind an opponent, the agent experienced reduced tire grip, resulting in missed corners or rear-end collisions that triggered reward penalties. Over thousands of hours of training, GT Sophy mapped the exact relationship between proximity to a leading car and the corresponding loss of braking and cornering grip. Consequently, it learned to tuck behind rivals on straightaways to build speed, pull out of the wake at the optimal moment to regain clean air and downforce for braking, and execute clean passes into the corners.