It has been discovered about a decade ago that autistic children cannot reason properly about mental states and mental actions of themselves and others (Baron-Cohen 1988, Leslie 1987). In particular, autistic patients demonstrate the lack of perceiving the emotions of themselves and others as well as the reduced emotional behavior. We extend the psychological observations and the "theory of mind" model by a more precise description of how the reasoning about mental states and actions are corrupted under autism. Logical artificial intelligence (AI) and, in particular, reasoning about knowledge, beliefs and action serves as a basis to construct a set of scenarios, consistently misunderstood by the autistic children in our experimental study. We build a set of exercises, teaching autistic patients to reason properly about mental states in accordance to the traditions of axiomatic method, since the natural ways of teaching usually do not help. Suggested autistic training is intended to stimulate the emotional development and to help the children to better communicate with external world. In this study, we continue the development of our logical model for the autistic reasoning (Galitsky 2000a) to build the training strategy of emotional rehabilitation. We provide theoretical and experimental evidence that the skills of formal reasoning about mental states and actions, acquired after the special training, lead to the improvement of emotional and communicational skills of the autistic patients. From the prospective of logical artificial intelligence, it is quite natural to approximate the emotional activity by reasoning about mental states and actions.