A model-free algorithm to safely approach the handling limit of an autonomous racecar

One of the key aspects in racing is the ability of the driver to find the handling limits of the vehicle to minimize the resulting lap time. Many approaches for raceline optimization assume the tire-road friction coefficient to be known. However, this neglects the fact that the ability of the system to realize such a race trajectory depends on complex interdependencies between the online trajectory planner, the control systems and the non-modelled uncertainties. In general, a high quality control system can approach the physical limit more reliable, as it applies less corrective actions. We present a model-free learning method to find the minimum achievable lap-time for a given controller using online adaption of a scale factor for the maximum longitudinal and lateral accelerations in the online trajectory planner. In contrast to existing concepts, our approach can be applied as an extension to already available planning and control algorithms instead of replacing them. We demonstrate reliable and safe operation for different vehicle setups in simulation and demonstrate that the algorithm works successfully on a full-size racecar.