Learning-Based Kinematic Control of a Deployable Manipulator With Long Span and Low Stiffness

Abstract

The deployable manipulator features long span and low stiffness during operation, resulting in significant positioning errors. The conventional mechanism control model based on error parameters is complex, lengthy, and challenging to guarantee accuracy in practice. This article proposes a learning-based kinematic control of the deployable manipulator. First, an analysis of kinematic performance and positioning error is presented. Then, the dataset is built by collecting data in the real environment, and the load factor that significantly influences the actual kinematics is taken as an extra feature. We propose a dataset building method based on manipulability according to the kinematic characteristics. A learning-based model consisting of a gated recurrent unit (GRU) and a 1-D convolutional layer is proposed, which is lighter and more effective than existing methods. Trajectory tracking and target grasping experiments are conducted to validate the performance of the kinematic control. The experimental results demonstrate that the proposed learning-based approach can achieve precise control under variable loads. This method could be extended to the kinematic control of similar deployable manipulators or flexible robots.