Learning By Demonstration for a Humanoid Robot Using Clustering

Robot skill acquisition can be conceptualized as a task of identifying patterns in a spatio-temporal sensory feature space. Skills can be mathematically learned through identifying mappings from sensory signals into Qualitative States (QS), construction of a QS skill automaton, and the detection of motor or output commands that transition the model from the present QS to the next QS. In this paper, the Fuzzy C-Means (FCM) and Robust Competitive Agglomerate (RCA) algorithms are used to identify QS clusters in the NASA Robonaut’s sensory feature space for a reach and grasp task. The high dimensionality of the Robonaut’s sensor feature space is addressed through Principle Component Analysis (PCA). Our findings are validated through similar results discovered when applying Hidden Markov Models (HMM), cluster validity, and episode extraction based on joint velocities. Index Terms – Clustering, Fuzzy C-Means, Robust Competitive Agglomerate, Skill Acquisition, and Learning by Demonstration.