ABSTRACT
Inspired by the stiffness adaptation of human arms when performing tasks, a framework of learning and generalization of variable impedance skills is proposed to help a robot manipulator acquire human-like skills and allows it to adapt to more complex task situations. EMG signals are used to estimate the stiffness of the human arm, which enables the robot to adapt the force profiles to different task situations with a natural regulation process. Several models such as dynamical movement primitives are utilized to generalize the learned skills to newly given task situations. Significantly, DMP is used to encode both movement trajectories and stiffness profiles under this framework. Furthermore, we combine the DMP with Gaussian Mixture Models to enable the robot to learn from a set of demonstrations of the same task, which enable robots to learn from multiple demonstrations and generate a better motion trajectory than that just from one demonstration. Finally, experimental studies on a Baxter robot are conducted to verify the approach, which proved the ability of our method in learning and generalizing multistep tasks.
