A new ARTMAP-based network is proposed, which is in part a generalization of Williamson's Gaussian ARTMAP. The training of the new network is based on activation and match functions that are equal and identical to Mahalanobis distance. The classification treats the clusters obtained through training as Gaussian mixture models. The training process has improved its efficiency due to the fact that the repeated searches for the resonant node have been eliminated. In addition, the inverse covariance matrices are computed recurrently. The new network is analyzed and compared with the fuzzy ARTMAP and Gaussian ARTMAP. The results from the new network have shown much better hit rates at fewer output nodes on several benchmark problems. A complexity analysis of the three networks is also provided.

Grasping objects with a multifingered hand in an enveloping manner, known as power grasp, has three major aspects: grasp planning, optimal contact force distribution and dynamic control of the desired contact forces. The focus of this paper is on the third problem. The maintenance of the desired optimal contact forces should be normally based on the complete contact force sensing, which is difficult to provide. The real, inexpensive force sensing resistor (FSR) type of touch sensors can measure only the normal component of the contact forces. This paper proposes a simple feedback controller based on such deficient feedback information.

Classification and discrimination of EMG patterns of preshaping hand motion are studied. The motions were performed by grasping six objects: small cylinder, large cylinder, small ball, large ball, small disk (precision, pinch grasp) and key (lateral grasp). An earlier work at SDSU has shown that the high successful classification rate of over 90% can be achieved. In this work the possibilities of hierarchical cluster discrimination and classification were explored. It was shown that hierarchical approach yields faster neural network training. It is also shown that the Kohonen's self-organizing maps can be successfully used in determination of the cluster hierarchy which characterizes a particular subject.

The common characteristic of existing parallel algorithms for evaluation of the robot inertia matrix on multiprocessor architectures is strict synchronization of processors, which limit the degree of the inherent parallelism. Such algorithms can, in the best case, deliver a time complexity of O(N), even with O(N/sup 2/) processors. Consideration is given to an approximate method in which all processors are decoupled, i.e., the processor synchronization, which handles the proper values of the propagated variables, is completely abandoned. Such a scheme drops the time complexity to O(N) for O(N) processors, and to O(1) for O(N/sup 2/) processors. Experimentation on the KUMARAN2 multiprocessor system based on MC68030/68882 processors, has shown surprisingly small approximation errors, which on average are below 1%. An analytical study of the approximation error is presented.>

An efficient parallel implementation of the robot inverse dynamics based on the recursive Newton-Euler formulation is considered. The algorithm basically partitions the computations related to a manipulator link among separate processors, resulting in a parallel architecture in which the number of processors equals the number of degrees of freedom of the manipulator. This has considerably low inherent parallelism if the forward propagation of velocities and accelerations and the backward propagation of forces and torques are synchronized. Therefore, in order to maximize the parallelism, the synchronization is completely removed, leaving each processor on the most recent values of the propagated variables. Since the mathematical approach for error analysis is too complex and the simulation approach is incomplete and unreliable, the algorithm is investigated on a real, multiprocessor machine, KUMARAN. A series of experiments with the model of the PUMA 560 manipulator and randomly generated inputs has shown surprisingly small errors. Some improvements of the algorithm which do not impair the time efficiency but provide further reduction of the approximation error are presented.>

Two approaches for the compensation of the position and orientation errors of robot manipulators due to variations of geometric parameters are considered. The first approach is based on the correction of the joint displacements obtained by using nominal inverse kinematics for desired position and orientation, while the second approach is based on redefining the desired position and orientation before applying the nominal inverse kinematics. It is shown that both approaches are equivalent in terms of compensation effect if the parameter variations are sufficiently small. However, the second approach does not require computation of the inverse Jacobian and can be used in the singular configurations of the manipulator. For computation of the position and orientation errors, which are required in both approaches, kinematic sensitivity functions are used. This eliminates the need for kinematic models, which considerably improves the time efficiency of the compensation algorithms. Closed-form expressions for kinematic compensation are derived for the PUMA manipulator.>

R-shell is an application-oriented extension of the regular UNIX shell that provides robotics-related data-manipulation and world-modeling capabilities. Its main purpose is to assist in the preparation of input data for the testing and evaluation of algorithms in the area of robotics. R-shell maintains a set of variables which can be defined and modified through a set of UNIX-like commands, and later accessed from independent user programs. R-shell commands generally fall into three groups: general commands, object commands and manipulator commands. General commands manipulate R-shell variables through various robotic-related and algebraic transformations. Object commands provide for the creation of a kinematic model of the robot environment. Manipulator commands define manipulator types, create manipulators, and perform various kinematic computations.>