Measuring motion complexity and its applications to learning of motion skills.
Abstract:
We introduce Motion Complexity as a measure indicating how much relatively meaningful motion instants are included in a motion trajectory. Relative meaningful motion instants are measured as motion significance by both temporal and spatial entropy of Gaussian mixture of motion data sets.
We firstly show that these two measures can classify human motion skills of a same domain like drawing from simple to complex one, which are well matched with our intuition. And then, we demonstrate that motion complexity is also used to determine which fitting task, for example among various type of peg-in-hole, is better for first training to master all of its related tasks by reinforcement learning.
Finally by introducing joint motion complexity, we show what motion instants need to be attended in human-human and human- object interaction.