Online learning-based feedforward-feedback control for path tracking of driving robot-vehicle

Abstract

Simplified dynamics models cannot accurately describe actual vehicle behavior under different road conditions, while complex models significantly reduce the computational efficiency of model predictive control (MPC). To obtain accurate vehicle dynamics models, this paper proposes an online Levenberg-Marquardt neural network (OLMNN) with a two-stage learning strategy. In the first stage, the OLMNN is trained offline using simulated data to capture the basic vehicle dynamics characteristics. In the second stage, an adaptive online Levenberg-Marquardt (OLM) algorithm is employed to collect real-time state data and update the network parameters in real time, enabling the model to adapt to nonlinear vehicle characteristics. A feedforward-feedback control strategy is designed to enhance computational efficiency. The feedforward control constructs an optimization problem based on the OLMNN to generate steady-state steering angles, while the feedback MPC focuses on correcting errors using a simplified predictive model. This control strategy avoids the computational burden of incorporating complex models into MPC while improving control accuracy. Simulations and hardware-in-the-loop (HIL) experiments validate that the proposed control strategy achieves excellent tracking performance and computational efficiency under various driving conditions.