Communication through Physical Interaction: Robot Assistants for the Elderly

This research work is a part of a broader research project which has the aim to build an effective and user friendly communication interface for assistive robots that can help the elderly to have an independent life at home. Such communication interface should incorporate multiple modalities of communication, since collaborative task-oriented human-human communication is inherently multimodal. For this purpose, data was collected from twenty collaborative task-oriented human-human communication sessions between a helper and an elderly person in a realistic setting (fully functional studio apartment). My research mainly focus on collecting physical interaction data in an unobtrusive way during human-human interaction and analyzing that data to determine how it can be implemented to communication interface for assistive robots particularly in elderly care domain. Thus a pressure sensors equipped data glove was developed. Based on the data collected from this glove, communication through physical interaction during collaborative manipulation of planar objects was studied. Subsequently, an algorithm was developed based on the laboratory data analysis which can classify four di erent stages of collaborative manipulation of planar object. This algorithm was later successfully validated on experiments performed in a realistic setting with subjects involved in performing activities of elderly care and determining human-human hand-over of planar object in real-time. Other than understanding the communication through physical interaction, this research also presents the methods for recognizing various physical manipulation actions that take place when an elderly is helped by a care-giver in cooking and setting of dinning table. This particular work was motivated by the natural language analysis of the data collected with helper and an elderly person which showed that the knowledge of such physical manipulation actions helps to improve communication through natural language. The physical interaction based classification methods are first developed through laboratory experiments and later successfully validated on the experiments performed in a realistic setting.