Dissertation

Trustworthy UAS: A Holistic Approach

Taylor, Max

2024

Overview

Unmanned Aerial Systems (UAS) are increasingly important. Farmers monitor crops and apply pesticides with UAS. First responders use UAS in applications ranging from fire fighting to search and rescue operations. There is potential for rapid shopping delivery by UAS. In all these applications, UAS work closely alongside humans. Onboard firmware controls the behavior of UAS. This dissertation studies ways to improve the quality of firmware. We start by presenting the first large-scale analysis of software defects (\"bugs\") reported in open-source UAS firmware. We examine nearly 300 reported bugs in the two most popular open-source systems (ArduPilot and PX4) and categorize the defects. Motivated by our findings, we propose three technologies to automate the detection and repair of UAS bugs. First, Avis seeks to automatically diagnose sensor bugs caused by misusing onboard sensors. Second, SA4U identifies unit type errors caused by incorrectly mixing values with different physical unit types (e.g., meters and minutes) in a computation. Finally, Scalpel automatically repairs bugs found by SA4U. Deep learning is increasingly used to provide advanced autonomous behavior for UAS. To support higher quality deep learning systems we propose checkd. Checkd automates checkpoint/restore policy configurations. Underlying checkd's contribution is the thesis that better tuned models yield better behavior. Checkd helps practitioners fine-tune models by reducing the overall cost to train.

• Publisher: ProQuest Dissertations & Theses
• Subject: Artificial intelligence / Computer Engineering / Computer science
• ISBN: 9798384089230