AI-enabled prediction of sim racing performance using telemetry data

Abstract

Despite the emerging and rapid progress of esports, approaches for ensuring high-quality analytics and training among professional and amateur esports teams are lacking. In this paper, we demonstrate how the application of data science techniques and Machine Learning (ML) approaches in esports, particularly in sim racing science, can illuminate the most important in-game metrics that dictate performance. Thus, using a professional racing simulator and MoTec i2 Pro (v1.1.5, Australia), we gathered extensive telemetry data from 174 participants, who completed 1327 laps on the Brands-Hatch circuit in the Assetto Corsa Competizione (v1.9, KUNOS Simulazioni). We clustered the obtained laps based on performance (lap-time), and then identified driving behaviors within performance groups. We also analyzed the feature subset obtained from a hybrid feature selection approach using two correlation analyses and three ML models.The best model achieved a prediction accuracy of 97.19%, demonstrating that the model effectively captured the critical factors that influenced driving performance during a lap. The results confirm that average speed is the most important metric, followed by lateral acceleration, steering angle, and lane deviation. Our analyses offer key metrics for refining training tools and techniques in sim racing performance improvement.

Keywords

• E-racing
• Artificial intelligence
• Machine learning
• Esport
• Simulated racing