Author ORCID Identifier

https://orcid.org/0009-0006-8969-5612

Semester

Fall

Date of Graduation

2025

Document Type

Thesis

Degree Type

MS

College

Statler College of Engineering and Mineral Resources

Department

Mechanical and Aerospace Engineering

Committee Chair

Nicholas S. Szczecinski

Committee Co-Chair

Gary Marsat

Committee Member

Sergiy Yakovenko

Abstract

Sensory inputs allow animals to perceive, react, and adapt to an environment. However, the sensory information received by the body, such as visual, auditory, and proprioceptive information, could become overwhelming, thus overloading the brain. Yet, animals can process all this information by canceling redundant signals from their surroundings, allowing them to be more sensitive to novel or unexpected signals in their environment. Each species (i.e., birds, fish, mammals) has its own way of using and filtering sensory information, from auditory to locomotion adaptivity. Cerebellar circuits contribute to sensory filtering in a variety of systems. In particular, research on Ghost knifefish has shown how cerebellar feedback generates a prediction of sensory inputs to filter out the redundant aspects of incoming sensory signals. These weakly electric fish generate and sense weak electric fields to detect objects and animals in their environment. Electro-sensory signals are passed from sensory neurons to cerebellar granular cells, which then feedback onto the primary sensory neurons. This feedback pathway enables the cancellation of redundant (i.e., predictable) sensory information, making downstream networks more sensitive to novel changes in input. The mechanism by which this cerebellar pathway generates predictive feedback is well characterized, and we will therefore replicate the functionality of this mechanism in a robotic arm simulation.

My goal is to engineer a model of this system that enables a simulated robotic arm to detect unexpected sensory stimuli related to its movement, so it could detect when it makes unintended contact in its environment. To accomplish this goal, I replicate the component of the fish’s cerebellar circuit. I modeled it as a “bank” or group of resonators, each with a unique resonant frequency. The input to the system, which is discerned as sinusoids, is interpreted as the sensory signal for the system. The ability of this system to generate a “prediction” to filter it out and let through unpredictable (novel) signals is based on the resonant properties, and it was therefore crucial to build the resonators with specific properties. Even though the system is meant to perceive the input signal as proprioception, I hypothesize that this mechanism is generalizable to other forms of sensory feedback, e.g., exteroception.

I show analytically that the network approximates the dynamics of a second-order linear oscillator, e.g., a spring-mass-damper system. Each resonator has a unique resonant frequency that can be tuned by changing parameters in the model, i.e., the capacitance of neuron compartments. The analytically derived resonant frequencies and damping ratios of the resonators were validated with simulations of individual resonators. To evaluate if the system could cancel predictable signals, which would be confirmed with a constant amplitude for the voltage response of the comparator, since any changes to the amplitude would indicate a novel signal, I first subjected it to signals at frequencies that the fish could encounter in the non-biotic portion of its environment (e.g., 2-32 Hz). The system was able to cancel predictable inputs, which were generated with the help of the output that was integrated with optimized weights to generate a signal that best filters out the sensory input by inhibiting the comparator with the previous cycle’s input, mediated by delayed synaptic transmission. When subjected to a “chirp” (brief increase in frequency), the system fails to filter it out, as expected. Then, the system was tuned to respond to more typical robot motion frequencies (e.g., 1-2 Hz). After tuning, I integrated these resonators into a MuJoCo simulation of a planar two-joint robot arm whose joint angles were fed into the bank of resonators. My network only fired when the arm encounters a perturbation, functioning as a dynamic perturbation detector. I discuss how to use this synthetic nervous system in robot control, including how this network could enable a robot to autonomously alter its mode of locomotion in response to persistent unexpected sensory signals.

Download

Share

COinS