A spatio-temporal extension to Isomap nonlinear dimension reduction
ICML '04 Proceedings of the twenty-first international conference on Machine learning
| Hi-index | 0.00 |
This dissertation addresses the problem of modularizing the capabilities of a humanoid agent into skill level behaviors. Our approach to this problem is to derive the skill level behaviors in a data-driven fashion from human demonstration. The humanoid agent is provided with a repertoire of basic skills by leveraging underlying behaviors in observed human movement. These skills serve as a foundation for endowing a humanoid agent with the ability to act autonomously. Given such a repertoire, a humanoid agent can autonomously perform functions such as control for various tasks, classification of human motion, and learning by imitation. Additionally, a repertoire of skills provides a common vocabulary for human-agent interaction and interface for non-technical users. We developed Performance-Derived Behavior Vocabularies (PDBV), an automated data-driven methodology for deriving a vocabulary of skill level behaviors from human motion data. PDBV assumes as input an unlabeled kinematic time-series of joint angle values, acquired from human performance demonstrative of multiple activities. We present spatio-temporal Isomap as an unsupervised dimension reduction technique for uncovering underlying spatio-temporal structure in kinematic motion. Using spatio-temporal Isomap, demonstrated motion data are clustered into groups of exemplars, where each group contains exemplars of an underlying primitive behavior. Exemplars of a behaviors can be generalized and realized as forward models that encode the nonlinear dynamics of the underlying behaviors in the joint angle space of the agent. Skills as forward models can be used in a variety of functions, including control and perception. We validate and evaluate the above approach to automated skill derivation in several ways. First, the methodology is empirically evaluated on multiple sources of time-series data, spanning scripted activities such as dancing and athletics, in order to validate input motion preprocessing, the structure of derived behavior vocabularies, realizing each behavior as forward models, and humanoid agent control. Second, we analyze and discuss our approach for deriving capabilities with respect to our empirical results. Lastly, we illustrate the utility of derived behavior vocabularies for use in movement imitation, addressing two subproblems: humanoid motion synthesis and human movement classification.