Networked Swarm Transformation and Disruption -- From Single to Combating Swarms
Abstract
In nature, autonomous swarms typically consist of individual agents with limited dynamics and simple rules, which collaborate to produce emergent large-scale behavior. How simple rules translate into self-organized dynamics is difficult to predict in general. Nevertheless, such rules dictate the emergent behaviors and controls produced, which create solutions to given tasks. We define such emergent solutions as swarm objectives, or intent. Emerging swarm objectives depend on many factors, including communication strength, noise, latency,and network topology among the interacting agents. Examples of robotic swarm objectives and emergent dynamics in defense applications are autonomous radar targeting, jamming, and shielding. How to design autonomous swarms with such emergent objectives is unknown, and will benefit significantly from prediction and control theory.Another central, and open problem in swarm dynamics involves multiple interacting swarms. We propose to develop theoretical principles describing how one swarm can capture another, change another~s pattern or divert its task, as a function of: agent detection methods (e.g., local sensing or communication through a network), the numbers of agents within each swarm, the relative motion between swarms, and the local rules defining inter and intra swarminteractions. Such a program has obvious relevance for defense related reconnaissance and combat. Our transformation theories of combating swarms will be informed and tested in mixed-reality experiments in collaboration with the University Pennsylvania.Finally, a theoretical foundation is lacking for how to predict the patterns and stability of a swarm from limited observations of its dynamics. We propose to train deep neural networks and reservoir computers on model and experimental swarming data, in order to classify the objectives of a swarm, infer its interaction rules, determine the preconditions for spontaneous transformations between different swarming patterns, and forecast longterm dynamics. Such a program will combine model-based analysis with machine learning techniques, which will be guided and validated with laboratory experiments. An exampleprescription might go as follows: one swarm is deployed to sense the motion of another, input these measurements into learning algorithms, infer the unknown swarm~s emergent objectives, and control the unknown swarm~s dynamics from our interacting-swarms theory.The results of our proposed work will be general principles and tools for disrupting and transforming multiple interacting swarms. Current mathematical theories, experiments, and data analysis are severely lacking in every aspect of this field of autonomy. As a result, we expect our approach to be a major leap in understanding, predicting and controlling the dynamical behaviors of interacting autonomous swarms.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Apr 25, 2019
- Source ID
- N000141912253
Entities
People
- Mong-ying Hsieh
Organizations
- Office of Naval Research
- United States Navy
- University of Pennsylvania