Home // ICAS 2020, The Sixteenth International Conference on Autonomic and Autonomous Systems // View article

Imitating Task-oriented Grasps from Human Demonstrations with a Low-DoF Gripper

Authors:
Timothy Patten
Markus Vincze

Keywords: Robotic grasping; task-oriented grasping; learning from demonstration; imitation learning; deep learning

Abstract:
Task-oriented or semantic grasping is important in robotics because it enables objects to be manipulated appropriately and used for their intended purpose. Many objects are human designed, therefore, we address the problem of learning task-oriented grasps by directly observing human behaviour. A person simply demonstrates the appropriate grasp, which is quick and convenient for any user in the real world. Our approach uses RGB images to track the object and hand pose, then employs a neural network to translate the human hand configuration to a robotic grasp with fewer degrees of freedom. Analysis shows that a variety of low-dimensional representations of the hand enable the mapping to be learned and that the model better generalises to new demonstrators handling new objects when the training data is augmented. Experiments with a mobile manipulator show that a robot successfully observes grasps and imitates the action on objects in various poses. This is accomplished immediately, without additional learning and is robust in real-world conditions.

Pages: 80 to 86

Copyright: Copyright (c) IARIA, 2020

Publication date: September 27, 2020

Published in: conference

ISSN: 2308-3913

ISBN: 978-1-61208-787-0

Location: Lisbon, Portugal

Dates: from September 27, 2020 to October 1, 2020