Master of Science in Electrical Engineering, University of Dayton, 2022, Electrical Engineering

Despite impressive advancements in the field of robotics, tasks such quick reaching movements, bipedal locomotion, and balance maintenance have shown to be a challenge. A possible reason for this is the predominance of feedback controls in robotics, which provide robust controllers at the expense of a slower response. The part of the human brain responsible for the performance of such tasks is the cerebellum, which functions exclusively in a feedforward way. Prior studies have shown cerebellum inspired controller's capabilities in movement learning, performing quick reaching movements, and functioning in uncertain environments. This thesis focuses on supervised learning cellular-level cerebellum computational models and its capability of performing balance related tasks. Through computer simulations, the innovative design was tested for the first time on the balancing of the inverted pendulum and double inverted pendulum. Another concept investigated in this work is the effect of cerebellum network size on performance, where among four different network sizes, the largest network ever simulated by the EDLUT spiking neural network simulator was created. Lastly, the controller's capability to transfer knowledge to another model performing the same task with different dynamics was evaluated. All controller sizes tested displayed impressive results on the inverted pendulum, quickly learning how to balance the pole. For the double inverted pendulum, all but the smaller sized network were able to achieve learning. The larger networks displayed better performance in both tasks, but the creation of even larger networks might be necessary to properly define the cerebellum network size effect on performance. The bio-inspired design was also shown to be capable of transferring knowledge, with an initially trained controller outperforming an initially naive controller on inverted pendulum models with different dynamics. The findings of this experiment show that cerebellum comp (open full item for complete abstract)

Committee: Raúl Ordóñez (Advisor); Terek Taha (Committee Member); Temesguen Kebede (Committee Member) Subjects: Electrical Engineering

Doctor of Philosophy, Case Western Reserve University, 2013, EMC - Mechanical Engineering

Robotic exploration is valuable in many practical applications. Even something as simple for a human as lawn-mowing is a nontrivial challenge for robotic controllers. It is possible to build a safe obstacle-edging reflex based on range sensor filtering. However, in some situations, a reflex is insufficient. Animals such as cockroaches depend on exploration, and the complexity of their strategies may inspire robotic approaches. After many hours of cockroach trials with and without goals (darkened shelters) and further experiments in analysis, we extracted a state-based algorithm that makes these types of decisions stochastically and also captures the shelter seeking bias in the path length of cockroaches. We call this algorithm RAMBLER, Randomized Algorithm Mimicking Biased Lone Exploration in Roaches. Further we find that this algorithm can be extended to predict behavior of cockroaches in arenas with clear barriers between the entrance and the goal. For robotics, this algorithm could add some variability to robot paths in situations where heuristics fail. For biology, this algorithm is a model that may help us better understand the decision-making process in the cockroach brain.

Committee: Roger Quinn (Advisor); Roy Ritzmann (Committee Member); Michael Branicky (Committee Member); Kiju Lee (Committee Member) Subjects: Biology; Computer Science; Engineering; Mechanical Engineering; Robotics; Robots