Real-time Motion Planning of Kinematically Redundant Manipulators Using Recurrent Neural Networks

Abstract

With the wide deployment of kinematically redundant manipulators in complex working environments, obstacle avoidance emerges as an important issue to be addressed in robot motion planning. In this chapter, the inverse kinematic control of redundant manipulators for obstacle avoidance task is formulated as a convex quadratic programming (QP) problem subject to equality and inequality constraints with time-varying parameters. Compared with our previous formulation, the new scheme is more favorable in the sense that it can yield better solutions for the control problem. To solve this time-varying QP problem in real time, a recently proposed recurrent neural network, called an improved dual neural network, is adopted, which has lower structural complexity compared with existing neural networks for solving this particular problem. Moreover, different from previous work in this line where the nearest points to the links on obstacles are often assumed to be known or given, we consider the case of obstacles with convex hull and formulate another time-varying QP problem to compute the critical points on the manipulator. Since this problem is not strictly convex, an existing recurrent neural network, called a general projection neural network, is applied for solving it. The effectiveness of the proposed approaches is demonstrated by simulation results based on the Mitsubishi PA10-7C manipulator.