Recurrent neural networks as kinematics estimator and controller for redundant manipulators subject to physical constraints

Abstract

Redundant manipulators could be efficient tools in industrial production as a result of their dexterity. However, existing kinematic control methods for redundant manipulators have two main disadvantages. On one hand, model uncertainties or unknown kinematic parameters may degrade the performance of existing model-based control methods subject to joint limits. On the other hand, existing model-free control methods ignore the existence of joint limits although they do not need to know kinematic models of redundant manipulators. In this paper, a quadratic programming (QP) scheme is elaborated to achieve the primary tracking control task of redundant manipulators as well as joint limits avoidance task. Besides, a gradient neurodynamics (GND) model is utilized to estimate the kinematics of redundant manipulators. Then, a primal dual neural network, which is employed to solve the QP problem, and the GND model are integrated towards developing a model-free control method constrained by joint angle and velocity limits for redundant manipulators. The visual sensory feedback is fed to the two neural networks. The efficacy of the proposed control method is demonstrated by extensive simulations and experiments, and the merits of the proposed method are also substantiated by comparisons.