Targeted spatio-temporal pattern formation and mimicry with autonomous multi-agent systems
Abstract
In nature, swarms consisting of agents with limited dynamics and simple rules of interaction can exhibit collective behaviors that effectively solve complex problems for the swarm. These emerging solutions are naturally robust, scaleable, and extensible; however applying a similar approach with synthetic agents and swarms has proven to be a challenging problem. We would like to specify simpleagent interactions such that the emergent dynamics are able to solve a given problem of practical interest, such as converging to agiven intensity or density distribution in space and time, or mimicking the dynamic patterns of a scalar field in the environment. Defense applications for these behaviors include camouflaging a robotic swarm, robust collective sensing, coverage, etc. Current approaches to these problems generally rely on central control algorithms or individually ``smart" agents. The proposed work addresses developing rules for autonomous swarms that can produce prescribed spatiotemporal patterns, given communication constraints and environmental noise.To make progress on the general problem of producing prescribed spatiotemporal patterns with coordinated and multi-agent teams, we propose to create models and methods that take as their basic ingredients simple, cooperative swarming agents with local sensing, controls, and mechanics. First, we will create new swarm-homotopy techniques where local controls are added to general physical models in order to continuously transform known emergent patterns into targeted ones. This approach will allow us to leverage our past work and produce swarm transformations with perturbative and smooth control. The main challenge here is to determine what local control forces, when mixed with pattern-formation rules, generate defined target patterns. To help determine parameters thatcontrol swarms to targeted patterns, we will use neural networks to calculate feedback-control signals that steer trajectories of continuous-time nonlinear dynamical systems on graphs defined by network topology, within the framework of neural ordinary differential equations (NODEs) and other related machine-learning techniques.Next, the proposed homotopy controls will be combined with simplepoint-process strategies where agents sense and share local density information between nearby neighbors, as well as target-densitymaps, and move to cooperatively converge to a target pattern. Unlike known methods, cooperative motion will be built-up systematically from local dynamics and communication only, and target-density stability analyzed using statistical mechanics. Third, we will develop swarm-mimicry approaches using cellular automata dynamics, where agents may be stationary but have direct control of their ``color" (or some general scalar field), and cooperatively learn to instantiate scalar-field patterns from their environment. This approach will aim to develop methods for designing instances of local cellular automata that are able to reconstruct a given sensed pattern by solving the inverse generative problem for cases more general than that of the well-known Turing patterns. To reduce dimensionality and facilitate the search for suitable pattern-formation rules and parameters, we envision leveraging deep variational autoencoder techniques as well as evolutionary optimization methods. Our proposed work is expected to advance the field of autonomy by creating novel methods that leverage simple and local dynamics for prescribed collective-pattern production. We expect to have an impact on artificial intelligence research, as well, since our work entails creating new hybrid machine-learning techniques that merge learning with many-agent dynamical systems control. Finally, the proposed research will aid in giving future warfighters capabilitiesto deploy many-agent teams and swarms to camouflage, spoof, hide high-value assets, and other battle-space tasks that require forming prescribed patterns.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 15, 2023
- Source ID
- N000142312434
Entities
People
- Klementyna Kasraie
Organizations
- Georgia Tech Applied Research Corporation
- Office of Naval Research
- United States Navy