Dexterous Manipulation Planning from Human Demonstration

In recent years, many multi-finger robotic hands have been designed and developed in order to emulate a human hand. The human hand has many sophisticated a lot of complicated dexterous skills that human themselves sometimes did not notice, for example an ability to grasp and manipulate tiny objects, a capability to regrasp an object without releasing and so on. To mimic or imitate these kind of dexterous skills to high degree-of-freedom mechanic hands, a planning is required and it is often referred as Dexterous Manipulation Planning (DMP). In this thesis, DMP is explored in context of learning object manipulation using programming by demonstration (PbD) framework. The idea of this paradigm to plan a dexterous manipulation by letting robot observes a stream of human demonstration. An advantage of PbD is that the planning discovered will also include human purpose in manipulation. This is corresponding to previous research which suggested that human has particular reasons on how they grasp or manipulate objects. This thesis proposes a novel method to segment a stream of human manipulative movement. This is a preliminary step to plan and mimic dexterous manipulation ability to robot hands in a PbD framework. Human manipulative movement is observed by a data glove, which provides information of 18 joint angles of human hand. Then a highly dimensional joint space is reduced to a lower dimensional space using principle component analysis. A segmentation is done based on two assumptions: First, a contact relation between hand and object would change when a coordinative movement of all joint angles of the hand changes it direction. Second, a coordinative movement of all joint angles are an approximately linear line in a reduced dimensional joint space. To test and apply the proposed method three manipulative movements of human is considered. Two of which are the repetition of same simultaneous movement. The other is more complex manipulation of a pen-like object called interdigital step. The results are compared and discussed with another previous research.