RoboFish: Mixed Shoals of Live Fish and Interactive Robots for the Analysis of Collective Behavior in Fish

点击翻译按钮获取中文摘要

Fish schooling is a popular model system for studying collective behavior. In a swarm, information on food or predators is distributed to all members without explicit communication channels through waves of movement changes, emanating from the animals that directly perceived the environmental cues. Understanding the underlying interaction mechanics elucidates collective group decision making and leadership processes. A variety of mathematical models have been proposed explaining upon which local cues individuals react and how this creates behavior on the collective level. These models work well in simplified computer simulations but fail to make specific predictions about the behavior of real-world systems. In contrast to fitting models to prerecorded behavior, we propose using interactive robots to validate, extend and learn models of collective motion in-loop with the natural system. Central to the proposal, we will use the Reinforcement Learning framework to identify optimal robotic behaviors in a leadership task. In real-time experiments, the robot will learn the appropriate interaction rules of live fish (guppies) by trial and error. This will deepen our understanding of the evolution of collective motion, will help clarifying its function and allow robust predictions of the system behavior. Our multi-robot platform and learning strategies can be applied to a variety of lab-based model systems. By opening the source code and hardware specifications we foster the development and dissemination of this young methodology.