Search papers, labs, and topics across Lattice.
This paper introduces a closed-loop autonomous raceline optimization framework using NURBS trajectories and CMA-ES, enhanced with controller-guided spatial feedback to adapt to track characteristics. The method leverages tracking errors as informative signals for a Kalman-inspired spatial update, constructing an adaptive constraint map for trajectory refinement. Experiments demonstrate a 17.38% lap time reduction in simulation and a 7.60% improvement on real hardware across varying tire friction levels, showcasing robustness.
Autonomous racecars can now learn to shave off significant lap time by "feeling" the track through controller feedback, adapting their trajectory on the fly.
We present a closed-loop framework for autonomous raceline optimization that combines NURBS-based trajectory representation, CMA-ES global trajectory optimization, and controller-guided spatial feedback. Instead of treating tracking errors as transient disturbances, our method exploits them as informative signals of local track characteristics via a Kalman-inspired spatial update. This enables the construction of an adaptive, acceleration-based constraint map that iteratively refines trajectories toward near-optimal performance under spatially varying track and vehicle behavior. In simulation, our approach achieves a 17.38% lap time reduction compared to a controller parametrized with maximum static acceleration. On real hardware, tested with different tire compounds ranging from high to low friction, we obtain a 7.60% lap time improvement without explicitly parametrizing friction. This demonstrates robustness to changing grip conditions in real-world scenarios.
