Samuel Zapolsky, Doctoral Dissertation Defense

February 13, 2017

Title: Statistical Rigid Body Mechanics
Date and Time: Friday, February 24th, 2017 at 10AM-12PM
Location: Science and Engineering Hall, Conference Room 4605
Adviser: Dr. Evan Drumwright
 

Abstract

The current approach to creating and improving robots and their control systems follows a cycle where complex mechanisms and controllers are iteratively designed, built, tested and then redesigned to address anomalous behaviors that were observed during in situ testing. During the initial cycles of this procedure, when unvetted control systems are tested on physical hardware, easily avoidable errors (e.g., unexpected collisions from inexact link geometry measurements or parasitic oscillation in actuators and passive elements) can have catastrophic consequences. Although testing robots in situ can be costly, often entirely new phenomena emerge from situated testing that were not observed throughout early computerized simulation-based testing. The discovery of such unanticipated behavior is currently considered a normal occurrence during robotics testing in situ. This occurrence of unanticipated behavior is due in part to the increased complexity of robotic systems (e.g., uneven terrain, impacting collision, unpredictable contact conditions) compared to typical automotive or aerospace applications that rely heavily on simulation-based testing before physical experimentation. I propose a statistical approach to simulation, where the indeterminacy of physical models or uncertainty in the structure of a mechanism or its environment is represented as a collection of particles in many parallel simulations. The aim of this approach is to inform roboticists of the possible unknown or unexpected behaviors that a robotic system may exhibit in order to address these faults before they affect the physical system. Our approach excites many of the errors caused by model, sensor, actuation, and communication error or uncertainty and can assist roboticists in determining whether certain robot designs, control systems, or modeling assumptions might result in unpredictable behavior. We propose using these detected faults toward designing more robust controllers and mechanisms.