Online Data-Driven System Identification and Adaptive Control Using Deep Koopman
Date:
Focus Area 1.1 - Virtual Prototyping of Autonomy-Enabled Ground Systems, real-time controls group
3 hour interactive poster presentation
Abstract: My research is focused on real-time data driven system identification and adaptive control using deep learning and Koopman spectral theory. Increasingly complex nonlinear dynamical systems are difficult to accurately model using classical dynamics, and existing linearization techniques are only locally valid. Koopman spectral theory enables us to accurately approximate transient dynamics of high dimensional nonlinear dynamical systems as linear dynamical systems. Our supervised deep learning framework enables us to learn from streaming sensor data, the best linearization of the system with minimal tracking error over long time horizons. We can perform system identification at high frequencies in real-time and then reformulate the optimal control law such that we track the desired trajectories of the system robust to changes in the dynamics of the system. For example, if the payload on a ground vehicle changes or if there are changes in the terrain, we can account for these changes in our adaptive model and reformulate the controller accordingly. This approach has been proven to exceed state of the art model predictive control algorithms for 1/10th scale race cars and we are going to implement our new algorithms onto a 1/12th scale tank in both simulation and on a physical system. Altogether this revolutionary perspective on dynamical systems and control theory will increase robustness and efficiency in autonomous ground vehicle systems for the army.