A control system framework for autonomous vehicles at the limits of handling

Autonomous Driving (AD) at the limits of handling represents a big technical challenge of which vehicle dynamics and control are at the core. The main complication is the thin line separating the vehicle from operating at its peak performance and underperforming. To evaluate the effect that the decision-making process has on vehicle performance, one can utilise path-following algorithms. In this research, a mathematical framework is developed which studies all areas of the vehicle dynamics and control of a car driving on a given trajectory in non-linear conditions. The framework comprises three main areas. First, the vehicle state estimation is studied, concentrating on the states which are not and cannot be measured on a vehicle with low-cost sensors, namely the sideslip angle. This can be observed with model-based methods or machine learning techniques. However, with model-based methods it is difficult to obtain accurate results given the large non-linearities and uncertainties in the dynamics and with machine learning techniques large datasets are required to avoid extrapolation. Therefore, an integrated Artificial Neural Network (ANN) and Unscented Kalman Filter (UKF) observer is developed which uses only Inertial Measurement Unit (IMU) measurements and can work as a standalone sensor. The goal is to use only numerical tools to develop the virtual sensor, therefore, the ANN is trained solely with numerical data obtained using a Vi-Grade model. The ANN outputs a pseudo-sideslip angle which is used as an input for the UKF. Since this is based on a kinematic model, the UKF is not affected by model mismatch and is capable of correcting the estimate despite model uncertainty. The UKF requires longitudinal velocity as an input, however, this is not measured directly by the IMU. Therefore, longitudinal velocity is estimated by means of direct integration with integral damping and integral reset value. The pseudo-sideslip angle is also corrected to improve the convergence of the UKF. Second, a robotic control scheme is developed for AD path-following at the limits of handling. This generally requires a compromise between computation time and model complexity. To tackle this a hierarchical controller made of two Non-linear Model Predictive Controls (NMPCs) is developed. The advantage of this type of scheme is that the two levels of the controller can interact to guarantee the desired outcome. While the higher level NMPC operates on a long prediction horizon, the lower level NMPC operates on a short horizon. The difference is that the high-level NMPC is based on a simple point-mass model and tyre-dependant “gg-diagram” constraint and is used solely to calculate velocity profiles. The output is then used as a terminal constraint and terminal cost by the lower level which is based on a seven degrees of freedom vehicle model with full Pacejka Magic Formula (MF) tyre formulation on all tyres, load transfers and Limited Slip Differential (LSD). Because of the precomputed terminal set, the low-level NMPC only requires a short horizon and focuses on exploiting the vehicle performance in real-time. For both controllers, the full Non-linear Optimisation Problem (NLP) is solved at each step. Third, the stability of the vehicle from a classic vehicle dynamics point of view is analysed with the aim of discussing how certain design paradigms can be changed when the human driver is replaced with a machine. Specifically, the vehicle stability of traditional vehicles is designed to be inherently stable leading to an understeering attitude. These design goals are required since it is safer for a human driver, however, this causes a decrease in peak lateral grip due to oversized rear tyres. With advanced autonomous driving features, the passive vehicle dynamics design goals could be set in a different way thereby gaining in both peak grip and transient response. Thus, several vehicle models with differing dynamic behaviours are developed and tested on a Driver-in-Motion (DiM) dynamic driving simulator driven by a number of expert human drivers. The same tests are then run again in a Model-in-the-Loop (MiL) simulation where the vehicle is controlled by means of a NMPC. The results pose some interesting questions on how commercial vehicles can be designed if driven by a robotic controller. The outcome of this work is a step further into the research of autonomous vehicles at the limits of handling from both an estimation and control perspective. This represents an advancement in the state of the art. The results of the framework developed show the effectiveness of the algorithms presented in this research at controlling a driverless vehicle and estimating its states by focusing on the vehicle dynamics.