Weakly-supervised traversability estimation for mobile robots using sparse point annotation

Schreiber, Andre Maurice

Permalink

https://hdl.handle.net/2142/121986 Copy

Description

Title

Weakly-supervised traversability estimation for mobile robots using sparse point annotation

Author(s)

Schreiber, Andre Maurice

Issue Date

2023-11-14

Director of Research (if dissertation) or Advisor (if thesis)

Driggs-Campbell, Katherine R

Department of Study

Electrical & Computer Eng

Discipline

Electrical & Computer Engr

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• mobile robots
• field robotics
• traversability
• weak supervision

Abstract

The task of traversability estimation is of key importance for mobile robots operating in unstructured environments. Such mobile robots need to be able to correctly predict where they can and cannot travel in order to ensure that they do not damage their environment or themselves. However, existing methods for estimating traversability in mobile robotics applications have significant limitations, as they can involve significant labeling tedium or require a robot to experience conditions of interest in order to label them. To overcome some of these challenges, we propose a weakly-supervised framework for traversability estimation in unstructured environments. Our framework uses sparse annotations to reduce labeling tedium, and we combine prior methods for weakly-supervised learning with a neural network-guided sampling strategy in order to improve results on our traversability prediction task. Our method demonstrates accuracy approaching that of the strongly-supervised approach, while significantly reducing labeling burden. Even though our initial proposed sparse annotation strategy demonstrates performance comparable to the strongly-supervised method with significantly reduced labeling effort, it can fail in cases of isolated, small obstacles. To address this issue, we propose a sampling strategy that explicitly emphasizes traversable regions and obstacles, and we demonstrate that this method can significantly improve prediction quality for small but untraversable obstacles.

Graduation Semester

2023-12

Type of Resource

Thesis

Copyright and License Information

Copyright 2023 Andre Maurice Schreiber

Owning Collections