This paper summarizes the online machine learning method LCore, which enables robots to learn to communicate with users from scratch through verbal and behavioral interaction in the physical world. LCore combines speech, visual, and tactile information obtained through the interaction, and enables robots to learn beliefs regarding speech units, words, the concepts of objects, motions, grammar, and pragmatic and communicative capabilities. The overall belief system is represented by a dynamic graphical model in an integrated way. Experimental results show that through a small, practical number of learning episodes with a user, the robot was eventually able to understand even fragmental and ambiguous utterances, respond to them with confirmation questions and/or actions, generate directive utterances, and answer questions, appropriately for the given situation. This paper discusses the importance of a developmental approach to realize personally and physically situated human-robot conversations.