Emergence of Specialization in a Swarm of Robots

We investigate the emergence of specialized groups in a swarm of robots, using a simplified version of the stick-pulling problem [5], where the basic task requires the collaboration of two robots in asymmetric roles. We expand our analytical model [4] and identify conditions for optimal performance for a swarm with any number of species. We then implement a distributed adaptation algorithm based on autonomous performance evaluation and parameter adjustment of individual agents. While this algorithm reliably reaches optimal performance, it leads to unbounded parameter distributions. Results are improved by the introduction of a direct parameter exchange mechanism between selected high- and low-performing agents. The emerging parameter distributions are bounded and fluctuate between tight unimodal and bimodal profiles. Both the unbounded optimal and the bounded bimodal distributions represent partitions of the swarm into two specialized groups.