MechE PhD Prospectus Defense: Mohammad Mehdi Kermanshah
- Starts: 10:00 am on Friday, October 25, 2024
- Ends: 12:00 pm on Friday, October 25, 2024
ABSTRACT: This PhD proposal presents a comprehensive approach to addressing un-certainty in autonomous systems through three levels. Firstly, we develop a filtering method to mitigate the effects of measurement uncertainty. Sec-ondly, we design a controller that remains robust in the face of measurement uncertainties, ensuring safety and stability despite imperfect measurements. Lastly, we employ a data-driven approach to estimate and model these un-certainties. The main contributions of each approach can be summarized as follows: Inspired by Complex Cell Networks (CCNs) in the primary visual cortex of mammals, we develop a novel filtering approach that emulates the biological system’s robustness and graceful degradation in the face of input deteriora-tion. Our formulation is founded on three principles observed in real neural responses: winner-take-all, persistence, and boundedness. We formalize these principles into mathematical constraints, and we derive model coefficients via Linear Programming (LP). These filters guarantee convergence for constant, bounded inputs and optimize both the convergence rate and sparsity of the filter kernel. We validate this approach by integrating the filter with a neu-ral network to estimate vehicle speed from camera images in extremely noisy environments. We propose a method to synthesize output linear feedback controllers for systems in polygonal environments with measurement uncertainties. We uti-lize perception modules that provide probability mass functions (PMFs) or rate maps capable of capturing nonlinear perception. Using this perception data, we formulate an optimization problem that incorporates Control Lya-punov Function (CLF) constraints to ensure stability and Control Barrier Function (CBF) constraints to ensure safety. Through robust optimization and the strong duality of LPs, we convert this problem into an LP that is solved offline. At a high level, our approach partially integrates perception, planning, and real-time control into a single design framework. We address the challenges of unmodeled dynamics in control systems by integrating data-driven modeling with model-based control. Specifically, we combine Model Predictive Control (MPC) with Gaussian Processes (GPs) as disturbance estimators for magnetically actuated cellbots. The GPs learn and predict unmodeled disturbances, and their predictions are utilized within the MPC framework to achieve precise trajectory tracking using relatively small datasets. Experimental validation demonstrates improved tracking ac-curacy, showcasing the effectiveness of merging data-driven and model-based techniques.
COMMITTEE: ADVISOR/CHAIR Professor Roberto Tron, ME/SE; ADVISOR/CHAIR Professor Calin Belta, ME/SE/ECE; Professor Sean Andersson, ME/SE; Professor Alyssa Pierson, ME/SE
- Location:
- PHO 901, 8 St. Mary's St.
- Hosting Professor
- Tron, Belta