A Framework for Automatic Behavior Generation in Multi-Function Swarms.

Area of Science:

• Robotics and Artificial Intelligence
• Multi-agent Systems
• Autonomous Systems

Background:

• Multi-function swarms aim to perform several tasks concurrently, presenting challenges in managing potentially conflicting requirements.
• Existing approaches often struggle with the dynamic adaptation needed for swarms operating in complex, multi-task environments.

Purpose of the Study:

• To propose a framework for automatic behavior generation in multi-function swarms capable of handling simultaneous, potentially conflicting tasks.
• To develop a system that allows swarms to dynamically transition between behaviors based on evolving situational requirements.
• To investigate the impact of noise and input variations on swarm controller performance.

Main Methods:

• Utilized the Quality-Diversity algorithm MAP-elites combined with a suitable controller structure for behavior generation.
• Tested the framework on a scenario involving simultaneous exploration, communication network creation, and Radio Frequency (RF) emitter geolocation.
• Employed ablation studies to assess the importance of individual sensor and controller inputs.

Main Results:

• Evolved a repertoire of controllers enabling swarms to transition between behaviors with varying task trade-offs.
• Found that moderate re-evaluations in MAP-elites enhance robustness with manageable computational costs.
• Demonstrated the impact of individual inputs on swarm controller performance through ablation.

Conclusions:

• The proposed MAP-elites based framework effectively generates adaptable behaviors for multi-function swarms.
• The evolved controller repertoire allows for online adaptation to dynamic task requirements and environmental conditions.
• Understanding input importance through ablation is crucial for optimizing swarm controller design and robustness.