Traditional tire modeling for autonomous racing requires extensive testing in controlled conditions - but what if we could learn tire behavior directly on the racetrack? We present a novel system identification algorithm that combines neural networks with traditional methods to learn accurate tire models during actual racing. Through iterative refinement and virtually generated data, our approach achieves 3.3x lower error rates than standard nonlinear least squares methods.
The key breakthrough: we can now learn a complete tire model with just 30 seconds of driving data and 3 seconds of training time, matching the accuracy of traditional steady-state methods - all while operating in dynamic racing conditions rather than requiring specialized testing facilities.
github: github.com/ForzaETH/On-Track-SysID
preprint: arxiv.org/pdf/2411.17508v1