Off-campus UMass Amherst users: To download campus access dissertations, please use the following link to log into our proxy server with your UMass Amherst user name and password.
Non-UMass Amherst users: Please talk to your librarian about requesting this dissertation through interlibrary loan.
Dissertations that have an embargo placed on them will not be available to anyone until the embargo expires.
Open Access Dissertation
Doctor of Philosophy (PhD)
Year Degree Awarded
Month Degree Awarded
Roderic A. Grupen
Erik G. Learned-Miller
Frank C. Sup IV
Artificial Intelligence and Robotics | Electro-Mechanical Systems | Robotics
Robots are increasingly expected to work in partially observable and unstructured environments. They need to select actions that exploit perceptual and motor resourcefulness to manage uncertainty based on the demands of the task and environment. The research in this dissertation makes two primary contributions. First, it develops a new concept in resourceful robot platforms called the UMass uBot and introduces the sixth and seventh in the uBot series. uBot-6 introduces multiple postural configurations that enable different modes of mobility and manipulation to meet the needs of a wide variety of tasks and environmental constraints. uBot-7 extends this with the use of series elastic actuators (SEAs) to improve manipulation capabilities and support safer operation around humans. The resourcefulness of these robots is complemented with a belief-space planning framework that enables task-driven action selection in the context of the partially observable environment. The framework uses a compact but expressive state representation based on object models. We extend an existing affordance-based object model, called an aspect transition graph (ATG), with geometric information. This enables object-centric modeling of features and actions, making the model much more expressive without increasing the complexity. A novel task representation enables the belief-space planner to perform general object-centric tasks ranging from recognition to manipulation of objects. The approach supports the efficient handling of multi-object scenes. The combination of the physical platform and the planning framework are evaluated in two novel, challenging, partially observable planning domains. The ARcube domain provides a large population of objects that are highly ambiguous. Objects can only be differentiated using multi-modal sensor information and manual interactions. In the dexterous mobility domain, a robot can employ multiple mobility modes to complete navigation tasks under a variety of possible environment constraints. The performance of the proposed approach is evaluated using experiments in simulation and on a real robot.
Ruiken, Dirk, "Belief-Space Planning for Resourceful Manipulation and Mobility" (2017). Doctoral Dissertations. 971.