Extended-Kalman-filter based parameter estimator applied to non-linear model predictive control of lateral dynamics of ground vehicles

Abstract: Research about model-based techniques (MPC) applications for automotive driving is nowadays experiencing an increasing interest thanks to improvements on hardware capabilities. Those allow for computationally heavy techniques as non-linear MPC is.

MPC-based techniques rely on models of the controlled plant for predicting future state-space vectors. Such models are either not accurate enough for the purpose or affected by uncertainties, hence potentially causing a lack of robustness. By updating the predictions model in run-time, learning-based MPC techniques are theoretically suitable to increase model/plant resemblance hence system performance.

This thesis is focused in the study of application of a learning technique when the uncertainty of the model is of type parametric, i.e., the structure and the order of the model is known, but some relevant parameters are uncertain orvarying. Such a learning technique is based on a recursion of the well-known extended Kalman filter, that makes use of the measured (noiseless) states for estimating the unknown parameters of the model.

The study is performed on a Matlab simulation environment where the lateral dynamics of a ground vehicle are to be controlled. for the purpose, different trajectories are used as references for a simulated vehicle that is intended to keep a minimum lateral deviation from them. Two different models are tested: kinematic and dynamic. For the first case, one parameter is unknown. It corresponds to the distance between the rear axis and the center of gravity (CoG) of the vehicle, which is not static in real situations. The dynamic model has two uncertain parameters, corresponding to cornering stiffness coefficients of both front and rear tyres. Those are depending on road-vehicle friction coefficient, hence could resemble an unknown road condition case.

Three main indicators are used to assess the performance of the system: lateral deviation, time and quality of parameter/s estimation. Simulation results proved that at the cost of an increased computational time, it is possible to improve the lateral dynamics as consequence of a better estimation value for the uncertain parameter.