We have been building a system called EMOTE to parameterize and modulate action performance. It is based on a human movement observation system called Laban Movement Analysis. EMOTE is not an action selector per se; it is used to modify the execution of a given behavior and thus change its movement qualities or character. The power of EMOTE arises from the relatively small number of parameters that control or affect a much larger set.

Our current project, sponsored by NSF, is to derive the Effort qualities from a live performance automatically via motion capture or video observation. Considering the possibility that a single motion may contain a few different phases in which different motion qualities are incorporated, we need to segment the motion trajectory of the end-effector into multiple appropriate phases. We plan to use a combined zero-crossing and curvature method to detect the descriptive changed in the motion. In our preliminary experiments, digital examination of 288 motion samples shows that the motion curvature is prominently high when the motion starts from rest, comes to a stop, or changes its direction, and that significant motion quality changes frequently arises at the turning points.

Because the neural network based inference system is trained and validated on the baselines of LMA notators, it is crucial to make the baseline motions as diversified as possible to cover different spatial directions, planes, dimensions, and have different forms. We plan to design and implement a choreography plan to include a comprehensive set of motion samples. Our preliminary experiments on 288 simple and short actions gave us a very positive result and encouraged us to extend our experiments to cover more complex, multi-segment actions.