*********************************
There is now a CONTENT FREEZE for Mercury while we switch to a new platform. It began on Friday, March 10 at 6pm and will end on Wednesday, March 15 at noon. No new content can be created during this time, but all material in the system as of the beginning of the freeze will be migrated to the new platform, including users and groups. Functionally the new site is identical to the old one. webteam@gatech.edu
*********************************
Title: Safe Robot Planning and Control Using Uncertainty-Aware Deep Learning
Committee:
Dr. Evangelos Theodrou - School of Aerospace Engineering
Dr. Ali-Agha - California Institute of Technology
Dr. Patricio Vela, Electrical and Computer Engineering
Dr. Samuel Coogan, Electrical and Computer Engineering
Dr. Fumin Zhang, Electrical and Computer Engineering
Abstract: In order for robots to autonomously operate in novel environments over extended periods of time, they must learn and adapt to changes in the dynamics of their motion and the environment. Neural networks have been shown to be a versatile and powerful tool for learning dynamics and semantic information. However, there is reluctance to deploy these methods on safety-critical or high-risk applications, since neural networks tend to be black-box function approximators. Therefore, there is a need for investigation into how these machine learning methods can be safely leveraged for learning-based controls, planning, and traversability. The aim of this thesis is to explore methods for both establishing safety guarantees as well as accurately quantifying risks when using deep neural networks for robot planning, especially in high-risk environments. First, we consider uncertainty-aware Bayesian Neural Networks for adaptive control, and introduce a method for guaranteeing safety under certain assumptions. Second, we investigate deep quantile regression learning methods for learning time-and-state varying uncertainties, which we use to perform trajectory optimization with Model Predictive Control. Third, we introduce a complete framework for risk-aware traversability and planning, which we use to enable safe exploration of extreme environments. Fourth, we again leverage deep quantile regression and establish a method for accurately learning the distribution of traversability risks in these environments, which can be used to create safety constraints for planning and control.
