The purpose of this project is the development of new algorithms for reinforcement learning (RL) problems in partially observable and noisy real-world environments. This is a typical situation found in designing effective behaviors policies in autonomous robots. The analysis of theoretically optimal behaviors and the use of consolidated machine learning techniques will be the basis for the development of the new algorithms. The unsupervised RL techniques are usually inefficient when applied to real-world environments. Therefore the new algorithms have to be trained and checked in supervised frameworks. In our approach, robustness and high learning rates will be achieved by using reward strategies that match the behavioral pattern structure to be learned. The basic structure of the trainer is independent from the specification of the agent and is applied to a specific context (tasks and environment) by a high level interface.