Due to a steady progress in interactive systems and robots, a natural evolution in the context of gaming experience is to bring the elimination of screens and devices for presenting the users with the possibility to physically interact with autonomous agents without the need to produce an entire virtual reality, such as that in classical videogames. This new style of game interaction, known as Physically Interactive Robogames (PIRG), exploits the real world (in both dynamical unstructured and structured aspect) as environment, and real, physical, autonomous entities as game companions. In this scenario, the present PhD research aims at investigating the use of machine learning techniques for developing complex, adaptive behavior in PIRG autonomous robots to the extent of supporting the development of on-line player modeling (which should also include an approach to intention detection) envisioning in-game behavior/strategy adjustment towards maintaining (or improving) the human player engagement. The planned methodology also aims to explore mobile robot bases with cheap sensors and algorithms requiring little power to be executed in real time ("green algorithms") in non-structured environments since these are constraints currently addressed in robogames and in the whole robotics community, to enable the spread of robots in the society and make them reach the market. As formal contribution to scientific community, the proposed research will open ways for the exploitation of new methods and approaches for designing PIRG in view of its relationship with ML-based techniques and Human-Robot Interaction. Moreover, it will add a new layer of exploration to the problem of creating playing robots even more able of being perceived as rational agents, i.e., possibly smart enough to be accepted as opponents or teammates, thus, becoming more likely to reach the mass-market as a new robotic product. Robogames are a well-suited framework to study free interaction between people and robots in a framed context: the methodologies and techniques developed in this work could also be used in other human-robot applications.