Julia Ebert, Harvard University
Coordinating behavior in a robot collective often requires the group to synthesize the beliefs of many individuals observing a large area into a single decision for the entire group. Such collective decisions are essential to allowing autonomous robot cooperation in applications from agriculture to space exploration. Extensive work has previously investigated the biologically inspired algorithms to solve this problem in a fully distributed collective, but such approaches often lack a mathematical grounding to allow interpretability and do not generalize beyond limited classes of decisions.
Here we present a Bayesian framework for collectives to make decisions about spatially distributed features of their environment with only local sensing and communication. This model allows tunable precision and decision points, depending on the application and acceptable speed/accuracy trade off. Each robot maintains an individual Bayesian model of a feature, updating its posterior from its own observations as well as those nearby robots communicate. When the concentration of the posterior-credible interval crosses a defined threshold, the robot then commits to a decision. We investigate this framework in the context of classification of a binary environmental feature at varying levels of simulated abstraction of the Kilobot robots, showing that a robot collective can accurately classify its environment. In addition, we demonstrate the benefit of a positive feedback mechanism to decision speed, similar to that observed in ant and honeybee colonies. Finally, we illustrate the generalizability of the Bayesian approach to a variety of spatial-distributed decisions (such as density and temperature), extensibility to multi-feature decisions, and potential benefits of informed movement and communication.
Abstract Author(s): Julia Ebert, Radhika Nagpal