Interpolative robot control with the nested network approach

A nested network method is presented for learning functions of high dimensions. The method, which is derived from the split-and-merge algorithm, creates a representation at multiple levels of coarseness from randomly distributed learning samples, and thus exhibits both fast and accurate learning. It is applied to learning the inverse kinematics in a three-degree-of-freedom pick-and-place problem. Without the need for building a model of the environment, the preprocessed sensor data is mapped onto joint displacements that must move the robot manipulator to the target object. Learning samples are obtained without a model of the manipulator. Instead the mapping from joint motion to camera motion is measured and taught directly to the nested network. A nested network method based on search trees adapts in real-time and reaches a grasping precision of up to 1-mm in only three steps