Robot skill learning through human demonstration and interaction

Abstract

Nowadays robots are increasingly involved in more complex and less structured tasks. Therefore, it is highly desirable to develop new approaches to fast robot skill acquisition. This research is aimed to develop an overall framework for robot skill learning through human demonstration and interaction. Through low-level demonstration and interaction with humans, the robot can learn basic skills. These basic skills are treated as primitive actions. In high-level learning, the complex skills demonstrated by the human can be automatically translated into skill scripts which are executed by the robot. This dissertation summarizes my major research activities in robot skill learning. First, a framework for Programming by Demonstration (PbD) with reinforcement learning for human-robot collaborative manipulation tasks is described. With this framework, the robot can learn low level skills such as collaborating with a human to lift a table successfully and efficiently. Second, to develop a high-level skill acquisition system, we explore the use of a 3D sensor to recognize human actions. A Kinect based action recognition system is implemented which considers both object/action dependencies and the sequential constraints. Third, we extend the action recognition framework by fusing information from multimodal sensors which can recognize fine assembly actions. Fourth, a Portable Assembly Demonstration (PAD) system is built which can automatically generate skill scripts from human demonstration. The skill script includes the object type, the tool, the action used, and the assembly state. Finally, the generated skill scripts are implemented by a dual-arm robot. The proposed framework was experimentally evaluated.