Graduate Thesis Or Dissertation
 

Adaptive Multiagent Traffic Management for Autonomous Robotic Systems

Public Deposited

Downloadable Content

Download PDF
https://ir.library.oregonstate.edu/concern/graduate_thesis_or_dissertations/hm50tx310

Descriptions

Attribute NameValues
Creator
Abstract
  • There is growing commercial interest in the use of unmanned aerial vehicles (UAVs) in urban environments, specifically for package delivery applications. However, the size, complexity and sheer numbers of expected UAVs makes conventional air traffic management that relies on human air traffic controllers infeasible. To enable UAVs to safely and efficiently operate in congested environments, it is essential to develop autonomous UAV management strategies. We introduce a dynamic hierarchical traffic control model that reacts to traffic conditions instantaneously to reduce congestion in the airspace. An obstacle-filled airspace lends itself to a modelling as a graph structure similar to a road network. We introduce controller agents, which set costs across the airspace. These agents control traffic similarly to adaptive metering lights in highway traffic. UAVs then plan their paths based on the costs (e.g. conflicts, or delays) they see for traversing particular parts of the airspace. This provides us a decentralized method for reducing traffic in an airspace Our hierarchical structure allows us to separate the traffic reduction problem from the individual robot navigation problem. Each robot does not explicitly coordinate with others in the airspace. Instead, robots execute their own individual internal cost-based planner to travel between locations. We then use neuro-evolution to provide incentives to these cost-based planners to reduce traffic in the environment. Traffic quality can be expressed in several different ways. We first evaluate traffic our traffic reduction policies in terms of `conflicts', which characterizes situations where an aircraft comes too close to another for safety in a physical space. We then examine traffic in terms of the amount of `delay' that all agents incur, which assumes that there is a structure to ensure only a safe number of UAVs occupy the same area. Finally, we look at the total travel time that a UAV can expect to take from the moment it enters the airspace until the time it gets to its destination. To facilitate an exploration of the UTM problem without waiting for a full simulation of UAVS running with A* , we develop an abstraction of the UTM domain that preserves the core UTM problem. We then investigate performance under differing levels of traffic, a well as two different agent structures. Our results show similar performance for both agent definitions, with delay reduction of up to 68% in high traffic cases. With a fast version of the UTM problem, we explore the effect of redefining the control structure such that links, or edges of the UTM graph, set costs individually. This shifts the control paradigm toward controlling directional travel rather than areas in the space, as was the case with sector agents used in previous approaches. Due to our graph structure, we find that there are far more control elements in the link agent approach than in the sector agent approach. We identify a tradeoff; link agents give finer control, but the coordination problem for the sector agents is easier because there are fewer sector agents. This indicates that we can improve performance out of a more distributed link-based setup if we address the challenges of multiagent coordination. However, the UAV traffic management domain presents a uniquely difficult coordination problem; each agent's action can affect the perceived value of every other agent's actions. This means that there is an excessive amount of noise in the system, as another agent's action can have a lot of impact on the reward an agent receives. We reduce the amount of multiagent noise by reducing the number of agents that are capable of learning. We identify that some agents have more ability to influence traffic based on the topology and traffic profile of the graph. This metric we call impactfulness. We use this metric to improve the learning by removing less impactful agents from the learning process, making a more stationary system in which the impactful agents can learn. The contributions of this work are to: - Introduce a cost-based traffic management approach that is platform-agnostic and fast to implement. - Develop a multiagent approach to setting costs in this traffic management system that is adaptive to traffic conditions and learns long-term effects of management decisions. - Create an abstraction of UAV traffic that captures key physical attributes, creating a fast and flexible simulation method. - Quantify agent contributions to system performance by experimenting with single agent learning, single agent exclusion, and a sliding number of agents learning in the system.
  • Keywords: Planning, UAV, Multiagent
License
Resource Type
Date Available
Date Issued
Degree Level
Degree Name
Degree Field
Degree Grantor
Commencement Year
Advisor
Committee Member
Academic Affiliation
Non-Academic Affiliation
Rights Statement
Publisher
Peer Reviewed
Language
Replaces

Relationships

Parents:

This work has no parents.

In Collection:

Items

Thumbnail Title Date Uploaded Visibility Actions
file details RebhuhnCarrieM2017.pdf 2017-11-08 Public Download