Title:
Barrier Functions and Model Free Safety With Applications to Fixed Wing Collision Avoidance
Barrier Functions and Model Free Safety With Applications to Fixed Wing Collision Avoidance
Author(s)
Squires, Eric G.
Advisor(s)
Egerstedt, Magnus B.
Editor(s)
Collections
Supplementary to
Permanent Link
Abstract
Robotics is now being applied to a diversity of real-world applications and in many areas such as industrial, medical, and mobile robotics, safety is a critical consideration for continued adoption. In this thesis we therefore investigate how to develop algorithms that improve the safety of autonomous systems using both a model-based and model-free framework. To begin, we make a variety of assumptions (e.g., that a model is known, there is a single safety constraint, there are no communication limits, and that the state can be sensed everywhere), and show how to guarantee the safety of the system. The contribution of the initial approach is a generalization of an existing method for creating a barrier function, which is a function similar to a Lyapunov function that can be used to make safety guarantees. We then investigate relaxing these initial assumptions. In some cases, new additional assumptions are required, performance may be reduced, or safety guarantees may no longer be available. We motivate the thesis with collision avoidance for fixed wing aircraft which can be viewed as a pairwise constraint on each pair of aircraft. This introduces the need for considering multiple safety factors simultaneously, and we show that an additional assumption is needed in this case. We then relax the assumption that the vehicles have unlimited communication and find that safety can still be guaranteed. However, it is possible in this case that the overriding safety controller may be more invasive than if more communication is allowed. When we then further relax the assumption that the state can be sensed at all times, safety can still be guaranteed in some specified situations but the system may be more permissive in approaching safety boundaries. We finally remove the assumption of a known model for dynamics. Although removing this assumption means the system is no longer guaranteed to be safe, the benefit is that it allows a safety designer to build a far less invasive override to get more performance out of the system.
Sponsor
Date Issued
2021-07-29
Extent
Resource Type
Text
Resource Subtype
Dissertation