Abstract
The robot Learning from Demonstration paradigm has proven to be an effective way in teaching a robot to learn from a demonstrator. However, it has been identified that there is a lack of practice in this field of research towards gettingrobots to learn tasks from animals, more specifically fish. In this thesis, we propose a novel Learning from Demonstration framework consisting of 6 modules aimed towards developing a fish-inspired robot controller such as 1) Task demonstration, 2) Tracking, 3) Analysis of fish trajectories, 4) Acquisition of robot training data, 5) Generating a perception-action robot controller, and 6) Performance evaluation. These modules provide a clear guide towards the goal of being able to have a robot learn from fish. We identified the challenges with tracking the fish and initially developed a user-friendly fish annotation program with a variety of useful features to allow digitizing multiple points along the fish
body and navigating throughout the video trials. The frames digitized went towards a new fish detection algorithm based on training an artificial deep neural network that can detect the fish throughout the video trials. Preliminary progress has also been made towards developing a new data analysis toolbox to aid with analysing fish trajectories to extract linear and angular velocities and study fish movements in relation to their environment. This research is a work in progress, but shows promising results and is a novel contribution to the field which will lead towards the development of a fish-inspired robot controller.
| Date of Award | 2023 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Otar Akanyeti (Supervisor), Fred Labrosse (Supervisor) & Patricia Shaw (Supervisor) |