Unsupervised Segmentation of Human Manipulation Movements Into Building Blocks

During the last years, new approaches were proposed in which robotic behavior is generated by imitating human movement examples. This process can be sustainably simplified by an automatic detection of the movement sequences which should be imitated. For this, automated approaches for human movement segmentation are needed to avoid time-intensive manual data analysis. Suitable examples for imitation learning are building block movements, which are basic movements that can be combined to solve different tasks. Recently, we introduced the velocity-based Multiple Change-point Inference (vMCI) algorithm, which automatically segments human demonstrations of manipulation movements into sequences with a bell-shaped velocity of the hand which is said to be a characteristic feature of manipulation building blocks. In this paper, the velocity of the hand as well as other features of human manipulation movements recorded with a marker-based motion tracking system are evaluated with respect to their suitability to detect segment boundaries of manipulation building blocks. Additionally, we perform a more intensive evaluation of vMCI compared to the original publication by evaluating the algorithm on different manipulation movement demonstrations recorded from several subjects and comparing the approach to other state-of-the-art segmentation algorithms. The results support the assumption that the velocity of the hand is one of the main features to detect segment boundaries in human manipulation movements and that the vMCI algorithm can detect these segment borders online and unsupervised, also in movement recordings with a noisy velocity.