In this work, we investigate the development of a real-time intelligent system allowing a robot to discover its surrounding world and to learn autonomously new knowledge about it by semantically interacting with humans. The learning is performed by observation and by interaction with a human. We describe the system in a general manner, and then we apply it to autonomous learning of objects and their colors. We provide experimental results both using simulated environments and implementing the approach on a humanoid robot in a real-world environment including every-day objects. We show that our approach allows a humanoid robot to learn without negative input and from a small number of samples.