In traditional evolutionary robotics, robot controllers are evolved in a separate design phase preceding actual deployment; we call this off-line evolution. Alternatively, robot controllers can evolve while the robots perform their proper tasks, during the actual operational phase; we call this on-line evolution. In this paper we describe three principal categories of on-line evolution for developing robot controllers (encapsulated, distributed, and hybrid), present an evolutionary algorithm belonging to the first category (the (μ + 1) on-line algorithm), and perform an extensive study of its behaviour. In particular, we use the Bonesa parameter tuning method to explore its parameter space. This delivers near-optimal settings for our algorithm in a number of tasks and, even more importantly, it offers profound insights into the impact of our algorithm's parameters and features. Our experimental analysis of (μ + 1) on-line shows that it seems preferable to try many alternative solutions and spend little effort on refining possibly faulty assessments; that there is no single combination of parameters that performs well on all problem instances and that the most influential parameter of this algorithm-and therefore the prime candidate for a control scheme-is the evaluation length τ. © 2012 Springer-Verlag Berlin Heidelberg.