Abstract

To realize effective heterogeneous multi-agent teams, we must be able to leverage individual agents’ relative strengths. Recent work has addressed this challenge by introducing trait-based task assignment approaches that exploit the agents’ relative advantages. These approaches, however, assume that the agents’ traits remain static. Indeed, in real-world scenarios, traits are likely to vary as agents execute tasks. In this paper, we present a transformation-based modeling framework to bridge the gap between state-of-the-art task assignment algorithms and the reality of dynamic traits. We define a transformation as a function that approximates dynamic traits with static traits based on a specific statistical measure. We define different candidate transformations, investigate their effects on different dynamic trait models, and the resulting task performance. Further, we propose a variance-based transformation as a general solution that approximates a variety of dynamic models, eliminating the need for hand specification. Finally, we demonstrate the benefits of reasoning about dynamic traits both in simulation and in a physical experiment involving the game of capture-the-flag.