Date of Award


Document Type


Degree Name

Doctor of Philosophy (PhD)


Mechanical Engineering

Committee Chair/Advisor

Phanindra Tallapragada

Committee Member

Javad Velni

Committee Member

Yue Wang

Committee Member

Ardalan Vahidi


Underwater robots are important for the construction and maintenance of underwater infrastructure, underwater resource extraction, and defense. However, they currently fall far behind biological swimmers such as fish in agility, efficiency, and sensing capabilities. As a result, mimicking the capabilities of biological swimmers has become an area of significant research interest. In this work, we focus specifically on improving the control and sensing capabilities of fish-like robots.

Our control work focuses on using the Chaplygin sleigh, a two-dimensional nonholonomic system which has been used to model fish-like swimming, as part of a curriculum to train a reinforcement learning agent to control a fish-like robot to track a prescribed path. The agent is first trained on the Chaplygin sleigh model, which is not an accurate model of the swimming robot but crucially has similar physics; having learned these physics, the agent is then trained on a simulated swimming robot, resulting in faster convergence compared to only training on the simulated swimming robot.

Our sensing work separately considers using kinematic data (proprioceptive sensing) and using surface pressure sensors. The effect of a swimming body's internal dynamics on proprioceptive sensing is investigated by collecting time series of kinematic data of both a flexible and rigid body in a water tunnel behind a moving obstacle performing different motions, and using machine learning to classify the motion of the upstream obstacle. This revealed that the flexible body could more effectively classify the motion of the obstacle, even if only one if its internal states is used.

We also consider the problem of using time series data from a `lateral line' of pressure sensors on a fish-like body to estimate the position of an upstream obstacle. Feature extraction from the pressure data is attempted with a state-of-the-art convolutional neural network (CNN), and this is compared with using the dominant modes of a Koopman operator constructed on the data as features. It is found that both sets of features achieve similar estimation performance using a dense neural network to perform the estimation. This highlights the potential of the Koopman modes as an interpretable alternative to CNNs for high-dimensional time series. This problem is also extended to inferring the time evolution of the flow field surrounding the body using the same surface measurements, which is performed by first estimating the dominant Koopman modes of the surrounding flow, and using those modes to perform a flow reconstruction. This strategy of mapping from surface to field modes is more interpretable than directly constructing a mapping of unsteady fluid states, and is found to be effective at reconstructing the flow. The sensing frameworks developed as a result of this work allow better awareness of obstacles and flow patterns, knowledge which can inform the generation of paths through the fluid that the developed controller can track, contributing to the autonomy of swimming robots in challenging environments.

Author ORCID Identifier




To view the content in your browser, please download Adobe Reader or, alternately,
you may Download the file to your hard drive.

NOTE: The latest versions of Adobe Reader do not support viewing PDF files within Firefox on Mac OS and if you are using a modern (Intel) Mac, there is no official plugin for viewing PDF files within the browser window.