Programming dexterous manipulation by demonstration

Current robot manipulation systems lack the flexibility and intelligence which are inherent in human manipulation of objects. Research in the field of dexterous manipulation has resulted in the development of multi-finger mechanical hands which can stably manipulate and regrasp objects without releasing. In addition, sensors have been added to the robot fingers which emulate the human sense of touch, including force, vibration and pressure distribution. Despite such progress, a robot is still a poor substitute in situations where it is too dangerous or expensive to place a human. The fundamental difference is the ability to adapt to new situations and respond to unexpected occurrences. This thesis examines three different approaches to this problem: autonomous manipulation, telemanipulation and programming by demonstration.; In autonomous manipulation, a robot hand is programmed to perform a specific task. A state diagram is used to set the control law for each particular phase of the task. The sensors on the fingers are used to detect events requiring a transition from one state to another. An extended example of a two-fingered robot hand manipulating an unknown object is used to display the capabilities and limitations of this approach.; In telemanipulation, a robot is directly controlled by a human user. In the current implementation, the user wears an instrumented glove which measures the joint angles of the hand. User hand motions are mapped into robot hand motions, allowing the user to remotely perform a task. In addition, the user may wear a haptic feedback device which can display the forces sensed by the robot. The performance results and analysis of human subjects performing prototypical tasks are presented.; In programming by demonstration, a layer of separation is created between the user and the robot hand. The user, wearing the instrumented glove, interacts with a virtual world through a virtual human hand. The motions of the virtual objects are tracked for the purpose of generating a control law for the robot to perform the same or similar motion to a physical object. The added layer robot motion while still maintaining human decision making. The issues and implementation of tracking intended human manipulation of a virtual object are discussed.