Driving policy distillation in autonomous racing with adaptive racing vocabulary and optimal driving guidance

Abstract

Autonomous racing poses unique challenges - including high-speed dynamics, close competition, and operation at the limits of vehicle performance - that are not fully addressed by current learning-based driving policies. In this paper, we propose a specialized neural network driving policy architecture for real-time autonomous racing to tackle these challenges. Our approach introduces an adaptive racing vocabulary that encodes track geometry and vehicle state information, enabling the policy to respond effectively to rapidly changing racing conditions. We further employ policy distillation with multiple cost heads guided by an optimal driving reference, thereby reducing the reliance on large expert-driving datasets. In addition, Bayesian optimization dynamically combines cost components (controllability, safety, speed, etc.), minimizing lap time while maintaining vehicle control. In high-fidelity vehicle dynamics simulations, the proposed architecture demonstrates robust and adaptive driving behavior, successfully handling the complex and demanding scenarios inherent in autonomous racing.