Cal Poly
Hallucination Techniques for Self-Supervised Synthetic Datasets for Mobile Robots
Abstract
dc:description.abstractClassical techniques in autonomous navigation struggle in tightly constrained spaces. Machine learning has been shown to perform better in these difficult environments but most techniques require large amounts of navigation experience for training. Using a new machine learning paradigm learning from hallucination (LfH), training data can be collected in a safe environment and not require supervision. Data is collected in real time while an agent performs a random walk in free space, supervision is not required as there are no obstacles for the robot to run into. After a random walk a post processing pipeline will hallucinate a safety corridor around the path the robot performed during its random walk. This will trick a learning policy into believing the labels are a result of avoiding an obstacle. Our experiments show that variations in the creation of the safety corridor can improve model performance. Our approach enables a robot to navigate previously unseen environments with real obstacles despite only being trained in a simulated environment, alleviating the difficulties of real world demonstrations.
Degree
thesis:*- Name thesis:degree_name
- MS in Engineering
- Discipline thesis:degree_discipline
- Electrical Engineering
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Colburn, Wyatt D
- Contributors dc:contributor
-
- Carlos Diaz Alvarenga
- Electrical Engineering
- College of Engineering
Subjects
dc:subject × 6Identifiers
dc:identifier.*- Repository record dc:identifier
- https://digitalcommons.calpoly.edu/theses/3216