Building large-scale robot systems: Distributed role assignment in dynamic, uncertain domains Alessandro Farinelli, Paul Scerri and Milind Tambe Dipartimento di Informatica e Sistemistica Univerista’ di Roma “La Sapienza” Via Salaria 113 CAP 00198 Rome, Italy Computer Science Department University of Southern California 941 W 37th Street, Los Angeles, CA 90089 farinell@dis.uniroma1.it, scerri@isi.edu, tambe@usc.edu Abstract For robot teams in large-scale, real-world domains, an effective approach for allocating roles to team members (role allocation) is critical. Unfortunately, role alloca- tion is extremely challenging in such real-world domains since robots face significant uncertainties in their own ca- pabilities and they may be faced with dynamic and con- tinuous changes in their capabilities. Previous work in multi-robotic and multiagent systems has provided algo- rithms for role allocation, but these algorithms often fail to explicitly represent and reason with uncertainty in a robot’s own capabilities, and often do not address dy- namic continuous changes in such capabilities. This pa- per presents two key contributions to address these lim- itations in large-scale settings. First, it presents a novel role allocation and re-allocation algorithm that explicitly reasons with uncertainty and dynamic changes in robot’s own capabilities. Second, it presents the application of a proxy-based architecture (previously applied only in agent-human settings) in large-scale robotic domains — demonstrating the application of a reusable infrastructure for multirobotic domains. Moreover this paper presents an experimental study based on simulations performed in order to evaluate the algorithm. These simulations pave the way to large real-world practical multi-robotic system involving 100s of robots. 1 Introduction Large numbers of robots are capable of achieving com- plex goals in distributed environments, provided they act together effectively. Teams of robots can achieve goals that individual robots or uncoordinated groups of robots cannot achieve. However, achieving effective teamwork in a dynamic, uncertain environment with robots that do not always have accurate models of their current situation and capabilities is very challenging. Algorithms for coor- dinating the robots must deal with the inherent distribu- tion of the team, unpredictable robot behavior uncertainty and dynamics in the environment. A key requirement for distributed team coordination al- gorithm is an effective approach for distributing the re- sponsibility for roles among the team members (role al- location). In this paper, we focus on real-world domains that provide a significant challenge for distributed role al- location due to three key properties. First, these domains are highly uncertain, and in particular, robots may face significant uncertainties in their own capabilities. This uncertainty goes beyond a low-level action (e.g., move- ment) uncertainty, and fundamentally relates to a signif- icant uncertainty in how well a robot can fulfill a role. Second, these domains are highly dynamic, where ex- ternal or internal events may cause significant dynamic changes to a robot’s capabilities. These changes are not just binary (from a fully functioning robot to a completely failed robot); rather, there may be a continuous dynamic degradation in robot capability over time. Third, these domains may involve a very large number (100) of robots. While there have been significant investigations of such role allocations in both the distributed robot and multi- agent literature, these are largely inadequate in address- ing a combination of the three challenges listed above. For instance, previous work in multiagent systems has provided algorithms for optimal initial role allocations based on capability analysis [14], combinatorial auction [6], distributed constraint optimization [8] and Belief De- sire and Intension [12]. Role reallocation has also been considered: team reorganization based on intends.that in the Shared Plan approach [5] or critical role failures in STEAM [13]. Within the distributed robotic literature, authors have proposed methods using auctions[16, 4], others utilize behavior based architectures[9, 15]. While these algorithms perform well for small teams, as appli- cations are scaled to larger and more dynamic and uncer-