Date of Award

Spring 2-28-2019

Document Type


Degree Name

Doctor of Philosophy (PhD)

First Advisor

Christoffer Heckman

Second Advisor

Nicolaus Correll

Third Advisor

Bradley Hayes

Fourth Advisor

Nisar Ahmed

Fifth Advisor

Mark Rentschler


Robust long term autonomy represents one of the most important targets for advancing robotic applications in the next 10 to 15 years, particularly because of the the advances in the socalled Classical Age of Simultaneous Localization and Mapping (SLAM) which place us in the age of Robust-Perception. Creating algorithms which allow robots to operate for years, unsupervised, in any environment is key step in the direction of true autonomy. This thesis presents a suite of algorithms to help enable robust long term autonomy, specifically robustness to a robot’s calibration parameters (internal knowledge the robot must possess in order operate) and to the environment the robot is situated in. Starting from the fundamentals of SLAM, the now de facto formulation is presented as a segue into self-calibration - the task of estimating calibration parameters such as the camera position on the robot. The following extensions are then developed: (i) an approach to treat slowly varying quantities, such as the position of a sensor drifting over years of operation, (ii) an algorithm which allows a robot to learn what movements it needs to perform in order to know its calibration parameters - using a reinforcement learning framework for self-calibration and (iii) all the insight from previous research is used to create a real-time self-calibration system which is capable of dealing with drift, unobservable parameters - for example when the robot is constrained to planar movement - and an information theoretic based segment selection mechanism which only choses “informative” segments of the trajectory in order to reduce computation time. However robustness is not only in regards to internal parameters such as a robot’s sensor position - the environment the robot operates in is dynamic - dealing with that environment is the final contribution, where an online probabilistic approximate joint feature persistence model is presented to determine which parts of the world are changing.

Available for download on Sunday, October 10, 2021