Track-Centric Iterative Learning for Global Trajectory Optimization in Autonomous Racing

Abstract

Minimum lap time trajectory planning is crucial for achieving competitive performance in autonomous racing, where real-time optimization must balance computational efficiency with trajectory optimality. However, existing approaches face significant challenges: long-horizon optimization is computationally intractable, and the inherent mismatch between analytical vehicle models and real-world dynamics limits performance. To address these limitations, this paper proposes a track-centric learning framework that directly optimizes a surrogate model of lap time rather than focusing solely on perfecting model fidelity. Our approach is twofold. First, we introduce a wavelet-based parameterization to represent trajectories in a low-dimensional space, making the optimization computationally tractable. Second, we employ Bayesian optimization to efficiently search this parameter space, where each candidate trajectory is evaluated by predicting its lap time through high-fidelity simulation incorporating a learned residual dynamics model. The effectiveness of the proposed framework is validated through simulations and real-world experiments, demonstrating lap time improvements of up to 17.6% over the nominal model. These results demonstrate that our surrogate-based optimization approach can effectively bridge the gap between model-based planning and real-world performance, offering a practical solution for high- performance autonomous racing applications.