Vision-based dynamical systems and learning

Nowadays, it is widely recognized that autonomous robots are “intelligent” machines, capable of performing complex tasks in a dynamical environment without explicit human intervention. Moreover, these tasks are performed while ensuring safe interactions between human and robots. Hence, ensuring Collision Avoidance is critical in designing any autonomous system. Concretely, most collision avoidance systems in mobile robots rely heavily on a computer vision machinery that provides the necessary feedback to the decision making and control units. In recent years, the advances in both machine learning algorithms and computer hardware have led to more wide application of deep neural networks in computer vision. The first part of this thesis focuses on providing methodologies for designing control strategies that guarantee collision avoidance for multi-agent systems without any information about the relative distances among the agents. The controllers are designed such that guarantee optimality of the designed activation function are provided. Finally, some examples are provided to illustrate the methodology, where some state-of-the-art activation functions are derived analytically they rely solely on the visual information, that is, times-to-collision and line-of-sight angle. These methods are particularly suitable to low-cost and/or small robotic systems that are not equipped with range measurement devices like radars and LIDARs. Furthermore, the collision avoidance is guaranteed using Lyapunov analysis type of technical arguments and illustrated using simulations. In the second part of this dissertation, a formal methodology for describing and designing activation functions in deep neural networks, is provided. The methodology is based on a precise characterization of the desired activation functions that satisfy particular criteria like circumventing vanishing or exploding gradient during training. The problem of finding desired activation functions is formulated as an infinite dimensional optimization problem, which is later relaxed to solving a partial differential equation. Furthermore, bounds that guarantee optimality of the designed activation function are provided. Finally, some examples are provided to illustrate the methodology, where some state-of-the-art activation functions are derived analytically.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Systems & Entrepreneurial Engr

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2022

Author dc:creator

• Amrouche, Massinissa

Contributors dc:contributor

• Stipanovic, Dusan
• Hovakimyan, Naira
• Sreenivas, Ramavarapu
• Beck, Carolyn

• Control Theory, Vision Collision Avoidance, Dynamical Systems, Machine Learning, Deep neural networks

Statement dc:rights

• Copyright 2021 Massinissa Amrouche

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/113161

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/113161