Koopman Operator for Modeling & Control of Autonomous Vehicles
04/17/2023
The path tracking control performance of an autonomous vehicle is crucially dependent upon modeling choices and subsequent system-identification updates. Traditionally, automotive engineering has built upon a range of increasing fidelity of white- and gray-box models coupled with system identification to garner greater explainability. In recent times, however, end-to-end deep neural-network based black box methods like deep imitation learning and deep reinforcement learning have come to showcase superior performance by implicitly capturing dynamics through data. While these methods provide for increased adaptability, they face challenges like explainability, generalizability, and sim2real gap. In this seminar, we highlight our research on the Koopman Operator based linear embedding of non-linear dynamics. The core idea stems from utilizing data-driven methods to capture the system dynamics in a “lifted” linear model allowing us to utilize the wealth of literature of linear system analysis for control design. Our work highlights the merger of data-driven and physically-motivated Extended Dynamic Mode Decomposition (EDMD) with deep-learning based autoencoder networks to capture discrete maps in lifted space. Finally, we briefly preview a range of modular, open-architecture, open-interface, and open-source hardware and software frameworks that are being developed/leveraged to support these efforts. These primarily include the AutoDRIVE Ecosystem and 1:10 scale F1TENTH cars along with potential deployments on the 1:5 scale Hunter SE and full-scale Autoware-enabled OpenCAV.
Speakers
Ajinkya Joglekar
PhD Candidate, CU-ICAR
Tanmay Samak
PhD Student, CU-ICAR
Chinmay Samak
PhD Student, CU-ICAR