A Learning Framework for Controlling Robotic Manipulators with Human-like Actions

Endowing robotic manipulators with human-like actions can make their behaviours more explainable and predictable, which improves the quality and safety of human-robot collaboration (HRC). A human-in-the-loop learning framework is proposed to learn the human-like posture features from human teleoperation demonstration. The learnt feature model can be used to generate human-like actions on the redundant manipulators. Moreover, an online relabeling approach is adopted in the training session to address the covariate shift problem, which reduces the total demonstration time to 10 min or less. The effectiveness of the human-like control method trained by the proposed framework is validated through comparing the posture between the human and the robot, and tracking the dynamic trajectories. A user study shows that, manipulators controlled with the proposed human-like posture constraint method are more friendly to the users in subjective feelings, and more acceptable as co-operative tools for non-professional users.