Inverse Recurrent Models - An Application Scenario For Many-Joint Robot Arm Control

This paper investigates inverse recurrent forward models for many-joint robot arm control. First, Recurrent Neural Networks (RNNs) are trained to predict arm poses. Due their recurrence the RNNs naturally match the repetitive character of computing kinematic forward chains. We demonstrate that the trained RNNs are well suited to gain inverse kinematics robustly and precisely using Back-Propagation Trough Time even for complex robot arms with up to 40 universal joints with 120 articulated degrees of freedom and under difficult conditions. The concept is additionally proven on a real robot arm. The presented results are promising and reveal a novel perspective to neural robotic control.