Carrying Heavy Objects by Multiple Manipulators with Self-Adapting Force Control

Abstract

Hybrid force/position control of multiple robot arms to jointly carry an object in the whole overlapping working space is a non-linear problem. Firstly, the control parameters change with each robot's conguration, and secondly, the dynamic models of the other participating robots are not available if the robots are controlled distributedly. We adopt an approach of increasing learning through practising { the way a human would do it. Based on a B-spline network, each robot learns to nd its optimal control parameters using on-line reinforcement signals. The robot controllers can exchange their learned parts of control surfaces in order to accelerate learning. We apply the approach to control three Puma-260 arms for jointly carrying relatively large and heavy objects. The successful experiments show that the automatic learning procedure converges in a short time and the resulting forces/torques can be reduced to the minimum. 1 Introduction This work aims at automatic develop...