Model predictive control is a very popular control scheme in a wide range of fields including driver assistance systems and autonomous robots. For example, in driver assistance systems, predictive control allows for improved safety and comfort. However, its implementation is a challenge in uncertain environments. Therefore, it is desirable to predict the evolution of the environment in which the controlled system operates. In other words, we pursue a highly accurate forecast of the environment so that we may achieve feasible and reliable action from the controller.

This dissertation presents a systematic framework that uses predictive control and forecasts of the future environment to operate under uncertainties and constraints. In particular, we focus on enhancing the performance of a predictive control scheme based on an accurate trajectory prediction of any targets controlled by humans (e.g., vehicles driven by human or humans themselves). We propose several motion-prediction-models using physics-based and data-driven approaches to improve the accuracy of the forecast. An interacting-multiple-model approach with Kalman filter techniques is useful in environments where it is difficult to have prior data sets such as disaster sites. Based on collected data from experimental vehicles, machine learning methods including hidden Markov models, convolution neural networks, and recurrent neural networks are used to enhance long-term predictions. Furthermore, we present predictive controls based on a probabilistic view of uncertain forecasts. The effectiveness of the proposed framework is demonstrated via applications such as human-companion robots, automotive adaptive cruise control, and autonomous lane change assist. The results of both simulation and real experimental data show the synergy between the motion prediction models and the predictive control designs.