Introduction Machine learning (ML) [24] is a means of automatically generating solutions that perform better than those that are hand-coded by human programmers. Such improvement is possible in problem domains where optimal solutions are di#cult to identify, i.e., when there are no models available that can accurately relate a system's dynamics to its performance. One such domain is the control of multi-robot systems. Mobile robots are notoriously di#cult to control in a robust, reliable, and repeatable fashion. The challenges stem from uncertainty inherent in physically embodied systems, including in sensors, e#ectors, and interactions between the system components and the environment. The behavior-based (BB) control paradigm [22, 1] provides a means of structuring robot controllers into collections of task-achieving modules or behaviors, such as exploration and obstacle avoidance. The modules operate in parallel and interact within the system and also through their e#ects on the en

1 Introduction and Motivation Multi-robot task allocation (MRTA) algorithms for heterogeneous groups ofrobots have to be able to differentiate between robots based on their performance in order to optimize allocation. Existing MRTA algorithms [?,?] gener-ally do this based on hand-coded information about the task utilities relative to each robot. Using hand-coded task utilities, however, these algorithms aretypically not sensitive to the effects of group dynamics, such as interference and synergy. These effects typically have to be estimated at runtime as theyare difficult to model due to their volatility and complexity.

We present an adaptive multi-robot task allocation algorithm based on vacancy chains, a resource distribution process common in animal and human societies.

We present an adaptive multi-robot task allocation algorithm based on vacancy chains, a resource distribution process common in animal and human societies. The algorithm uses individual reinforcement learning of task utilities and relies on the specializing abilities of the members of the group to promote dedicated optimal allocation patterns. We demonstrate through experiments in simulation, the difference between the allocation patterns emerging when robots used greedy and softmax action selection functions. We conclude that using softmax functions makes the vacancy chain algorithm sensitive to different levels of ability in a group of heterogeneous robots as well as to the effects of the underlying group dynamics such as interference and synergy.