A computational framework for integrating robotic exploration and human demonstration in imitation learning

This paper proposes a computational framework for humanoid robots to learn complex behaviors through combining robotic self-exploration and demonstrations of humans. A modified Rapidly-growing Random Tree (RRT)-Connect algorithm is used for exploration, a Linear Global Model (LGM) is used for recording demonstrations, a spatial-temporal extension of Isomap algorithm is used for dimension reduction which enables the exploration in a low-dimensional latent space, and the log likelihood function of the distribution of sampled data in the joint space is used to project the data in the latent space back to the joint space. An experiment of imitating a conducting behavior is carried out to demonstrate the effectiveness of this framework.