An algorithm has been developed to solve the problem of mapping from (1) a glove instrumented with joint-angle sensors to (2) an anthropomorphic robot hand. Such a mapping is needed to generate control signals to make the robot hand mimic the configuration of the hand of a human attempting to control the robot. The mapping problem is complicated by uncertainties in sensor locations caused by variations in sizes and shapes of hands and variations in the fit of the glove. The present mapping algorithm is robust in the face of these uncertainties, largely because it includes a calibration subalgorithm that inherently adapts the mapping to the specific hand and glove, without need for measuring the hand and without regard for goodness of fit.
The algorithm utilizes a forward-kinematics model of the glove derived from documentation provided by the manufacturer of the glove. In this case, “forward-kinematics model” signifies a mathematical model of the glove fingertip positions as functions of the sensor readings. More specifically, given the sensor readings, the forwardkinematics model calculates the glove fingertip positions in a Cartesian reference frame nominally attached to the palm.
The algorithm also utilizes an inversekinematics model of the robot hand. In this case, “inverse-kinematics model” signifies a mathematical model of the robot fingerjoint angles as functions of the robot fingertip positions. Again, more specifically, the inverse-kinematics model calculates the finger-joint commands needed to place the fingertips at specified positions in a Cartesian reference frame that is attached to the palm of the robot hand and that nominally corresponds to the Cartesian reference frame attached to the palm of the glove.
Initially, because of the aforementioned uncertainties, the glove fingertip positions calculated by the forward-kinematics model in the glove Cartesian reference frame cannot be expected to match the robot fingertip positions in the robot-hand Cartesian reference frame. A calibration must be performed to make the glove and robot-hand fingertip positions correspond more precisely. The calibration procedure involves a few simple hand poses designed to provide well-defined fingertip positions. One of the poses is a fist. In each of the other poses, a finger touches the thumb. The calibration subalgorithm uses the sensor readings from these poses to modify the kinematical models to make the two sets of fingertip positions agree more closely.
In tests of software that implements the algorithm, the entire calibration process was found to take less than 30 seconds. Operators immediately noted a difference between the accuracy of fingertip positions as computed by this algorithm and as computed by a prior algorithm. The increased accuracy afforded by this algorithm was found to improve control of the robot hand. The algorithm and software were also adapted to use with an optically tracked glove for hand control, with similar results.
This work was done by Michael Goza of Johnson Space Center. For further information, contact the Johnson Commercial Technology Office at (281) 483-3809. MSC-23680