Robot Learning Assembly Tasks from Human Demonstrations

The industry robots are widely deployed in the assembly and production lines as they are efficient in performing highly repetitive tasks. They are mainly position-controlled and pre-programmed to work in well-structured environments. However, they cannot deal with dynamical changes and unexpected events in their operations as they do not have sufficient sensing and learning capabilities. It remains a big challenge for robotic assembly operations to be conducted in unstructured environments today. This thesis research focuses on the development of robot learning from demonstration (LfD) for the robotic assembly task by using visual teaching. Firstly, the human kinesthetic teaching method is adopted for robot to learn an effective grasping skill in unstructured environment. During this teaching process, the robot learns the object's SIFT feature and grasping pose from human demonstrations. Secondly, a novel skeleton-joint mapping framework is proposed for robot learning from human demonstrations. The mapping algorithm transfers the human motion from the human joint space to the robot motor space so that the robot can be taught intuitively in a remote place. Thirdly, a novel visual-mapping demonstration framework is built for robot learning assembly tasks, in which, the demonstrator is able to teach the robot with feedback in real-time. Gaussian Mixture Model and Gaussian Mixture Regression are used to encode the learned skills for the robot. Finally, The effectiveness of the approach is evaluated with practical assembly tasks by the Baxter robot. The significance of this thesis research is on its comprehensive insight of robot learning from demonstration for assembly tasks. The proposed LfD paradigm has the potential to effectively transfer human skills to robots both in industrial and domestic environments. It paves the way for general public to use the robots without the need of programming skills.