Configuration Depending Crosstalk Torque Calibration for Robotic Manipulators with Deep Neural Regression Models

Abstract

In this paper, an approach for articulated robotic manipulator which minimizes configuration depending crosstalk torques is presented. In particular, these crosstalk torques are an issue for the Kinova Jaco 2 manipulator. We can experimentally show that the presented approach leads to crosstalk minimization for the Kinova Jaco 2 manipulator. Crosstalk leads to a significant difference between sensor output and inverse dynamic models using CAD rigid body parameters. As a consequence, these disturbances lead to a hindered torque control and perception. Different machine learning techniques, namely Random Forests and various neural network architectures, are evaluated on this task. We show that particularly deep neural regression networks are able to learn the influence of the cross torques which improves perception.

BibTeX

@inproceedings{ZwienerIAS2018,
  title = {Configuration Depending Crosstalk Torque
Calibration for Robotic Manipulators with Deep
Neural Regression Models},
  author = {Adrian Zwiener and Sebastian Otte and Richard Hanten and Andreas Zell},
  booktitle = {Intelligent Autonomous Systems (IAS), The 15th International Conference on},
  address = {Baden-Baden, Germany},
  year = {2018},
  month = june,
  note = {(Accepted for publication)},
  abstract = {In this paper, an approach for articulated robotic manipulator which minimizes configuration depending crosstalk torques is presented. In particular, these crosstalk torques are an issue for the Kinova
Jaco 2 manipulator. We can experimentally show that the presented
approach leads to crosstalk minimization for the Kinova Jaco 2 manipulator. Crosstalk leads to a significant difference between sensor output
and inverse dynamic models using CAD rigid body parameters. As a consequence, these disturbances lead to a hindered torque control and perception. Different machine learning techniques, namely Random Forests
and various neural network architectures, are evaluated on this task. We
show that particularly deep neural regression networks are able to learn
the influence of the cross torques which improves perception.},
}