In this paper, we demonstrate a multi-phase genetic programming (MPGP) approach to an autonomous robot learning task, where a sumo wrestling robot is required to execute specialized pushing maneuvers in response to different opponents' postures. The sumo robot used has a very simple, minimalist hardware configuration. This example differs from the earlier studies in evolutionary robotics in that the former is carried out on-line during the performance of a robot, whereas the latter is concerned with the evolution of a controller in a simulated environment based on extended genetic algorithms. As illustrated in several sumo maneuver learning experiments, strategic maneuvers with respect to some possible changes in the shape and size of an opponent can readily emerge from the on-line MPGP learning sessions.