Precision manipulation with a dextrous robot hand

Despite the building of a number of complex robot hands in the last two decades, there have been few experimental demonstrations of robot hand dexterity. This paper outlines several key issues associated with precision manipulation for robot hands. \Precision ma-nipulation" is deened as the control of grasped object position and applied force using nger-tip contacts alone. The approach taken is to deene a set of skills|primitive manipulation functions|which the robot can perform. The manipulations are generalizable in the sense that they take parameters for diierent object geometries, speed and direction of motion. A single manipulation can also be performed with a number of diierent grasp topologies. With each manipulation there are several associated computations. First, the trajectories of the contact points can be calculated a priori from knowledge of the desired object motion. Second, a workspace analysis is performed to determine that the manipulation is within the workspaces of all of the ngers simultaneously and also to calculate the initial grasp positions and hand/wrist position in a global coordinate frame. Third, a task partitioning is performed to specify force-and position-controlled directions of the contact points. This partitioning controls the grasping forces on the object during the grasping and manipulation phases. The paper also describes how the primitive manipulations can be combined into complex tasks and several examples are presented, particularly removing a top from a child-proof bottle and rotating an object within the grasp. The basic and complex manipulations were implemented on a Utah/MIT dextrous robot hand system and experimental results are presented for all tasks. A hybrid force/position controller was also built to enable the experiments and it is brieey described in the paper as well.