The apparently simple act of moving a robot manipulator so as to grasp a target object can involve many complex issues, particularly when the position of the object is imperfectly known.
The research described in this paper considers a number of interacting issues that complicate this task. First, the position of the object is known only imperfectly; a probability distribution over a range is given, presumably obtained from a vision system. New information is gathered as a result of moving the arm toward the object, either by finding a contact or by not finding a contact. Second, collisions between the arm and the object may result in moving the object; this can either happen accidentally, or it can be the result of a deliberate attempt to move the object. Third, movements are considered with three kinds of goals: to actually achieve the grasp; to rotate the object into a position where the desired grasp is feasible; or to gather information about the position of the object.
The technique for planning these motions uses a decision-theoretic framework. A belief state records a probability distribution for the position of the object; this guides the choice of motion, and is updated as more information is acquired. A library of motions, described in terms of the object’s frame of reference, of each of the three categories of goals, is compiled offline, partly by manual training and partly by simulation.
The combinatorics both of belief update and of choosing a proper action soon become daunting; the major technical issue addressed in this research is the development of strategies to keep these manageable.
The system has been extensively tested on a simulated robot, and to a limited degree on a real robot, and has achieved an impressive degree of success. It is a substantial contribution to our understanding of techniques for planning robotic manipulations.