Dynamic Data and Information Processing
Abstract
This project aims to develop the new fundamental knowledge, methods and technological solutions required to enable learning-based verifiably safe navigation of aerial robotic systems operating inside geometrically-complex, sensor-degraded and contested environments. In particular, a novel navigation policy, derived through assured reinforcement learning techniques, is proposed that will guide an autonomous flying system across reference waypoints, while autonomously ensuring a) the avoidance of 3D obstacles and “no-go” zones inside dense and contested environments, b) sampling of data from the subsets of the environment that present interest to the robot’s mission, and c) accounting for the effects of localization uncertainty to the problem of safe navigation. To offer guarantees for field deployment, the envisioned work aims to introduce reachability analysis inside the training process and a verification tool to certify the safety and performance of the learned policy. The results will be demonstrated in the context of challenging field experiments involving low-altitude high-speed flight and maneuvering through canyons and fjords across Norway. Aiming towards challenging field conditions, we target environments such as Jutulhogget, the largest canyon in Northern Europe, as well as Trondheim’s fjord, alongside more northern ice settings.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 11, 2022
- Source ID
- FA86552117033XX0
Entities
People
- Kostas Alexis
Organizations
- Air Force Office of Scientific Research
- Norwegian University of Science and Technology
- United States Air Force