Modelling and generalizing achieved robot skills with temporal Restricted Boltzmann Machines

Abstract

To equip a robot with various required skills so that to serve human society, plenty of research have been performed and successfully applied from both theoretical and practical aspects for decades. Usually, a robot with several skills needs to recall a different controller or model parameters to fit the new circumstance as task to be fulfilled or environment changes. Therefore, how to smoothly shift to a suitable control model to handle changed circumstance becomes an important issue in robotics field. It will be more remarkable when robot skill(s) is extremely environmental sensitive (the skill must be modeled by a specific controller parameter(s) under given certain environment), or when different robot skill is modeled by completely different controller. In this research, a new approach is proposed to model achieved robot skills uniformly, where the modelling process is completed under a unified framework by employing Temporal Restricted Boltzmann Machines (TRBMs), and thus the shifting between different skills becomes more smoothly and easily. In achieving a robot skill, expertise hand tuning and autonomous learning are two commonly used styles, however both of them are still suffer from either boring timeconsuming or high learning computational complexity. As a consequence, how to efficiently achieve a new robot skill is an all-the-time issue to be tackled. In this paper, take the view of utilizing the knowledge that is inherent in those already achieved skills, the generalization of the proposed unified model is investigated. Experimental results on a humanoid robot PKU-HR5.1 suggest that the proposed approach is effective in both modeling and generalizing the achieved robot skills.

Learning Actions to Improve the Perceptual Anchoring of Objects

“…They have been previously used for tasks such as modeling full-body human motion (Taylor et al, 2007), learning motion primitives from demonstration (Kulić and Nakamura, 2011), style-content separation and motion style interpolation (Chiu and Marsella, 2011), and recently modeling robot walking motion under varying circumstances (Luo et al, 2014).…”