Vision-based Robot Handwriting Skill Reproduction and Generalization

Robot skills transfer can increase the availability and acceptability of robots. However, the physical contact between robots and humans when teaching is unsafe. Teaching based on teleoperation can avoid contact, but the requirement of the extra input device is inconvenient. In this paper, a possible solution is provided to address the problems by proposing a novel robot skill learning framework to realize handwriting skill transfer and generalization. Instead of teaching hand-to-hand, the computer vision technique is utilized to recognize the human motion in the demonstration stage, which makes skill transfer more intuitive and natural. This approach detects continuous curves by finding the connection point in pictures written by humans, and then the continuous curves are used for teaching so that robots can learn the human skills and further generalize the learned handwriting skills. Finally, an experiment that the robot writes the Chinese character "Yang" is conducted to validate the effectiveness of the presented approach.