We introduce the multi-model parameterized Koopman (MMPK) framework, a novel end-to-end data-driven modeling and control pipeline for enabling autonomous navigation in Uncrewed Ground Vehicles. MMPK builds upon the Koopman extended dynamic mode decomposition (KEDMD) algorithm, offering a flexible model- and control-adaptation in the presence of time-varying uncertainties with both ego-vehicle and operational-environment parameters. Unlike traditional methods, MMPK addresses challenges such as overfitting and reliance on a singular global model by adopting a set of pose-agnostic representations of positional data and curvature-parameterized Koopman models, thereby effectively mitigating data bias. The end-to-end unified pipeline encompasses: (i) an offline data-driven learning phase to customize the multiple curvature-parameterized Koopman models and (ii) an online model-based trajectory planning and linear Model Predictive Control (outer-loop control design) adapted to switched Koopman dynamics. The performance of the proposed pipeline is verified via simulation and experimental testing using a 1 / 5 th scale Ackermann-steered ground vehicle platform (AgileX Hunter SE) and benchmark driving profiles. Comparative evaluations demonstrate MMPK’s superior path-tracking capabilities and the effectiveness of its local planning strategy in bridging the Model-Sim-Real gap.
Joglekar et al. (Mon,) studied this question.