Learning Reduced-Dimension Models of Human Actions

Abstract

A great deal of current robotics research studies the modeling of human reaction skills: learning control mappings to represent a person's task-performance strategies. The important related problem of modeling human action skills has received less attention. Action learning is the characterization of the state space or action space explored during typical human performances of a given task. Action models learned from human performances typically represent some form of prototypical performance, and also characterize how the human's performances vary stochastically or due to external influences. They are used for gesture recognition, for realistic computer animations of human motion, for study of an expert performer's motion (e.g., Tiger Woods' golf swing), for generating feed-forward or reference robot control signals, and for evaluating the naturalness of the performances generated in real and simulated systems by reaction-skill models. This thesis formulates the process of building action models from human performance data as a dimension reduction problem. Characterizing action skills involves determining the lower-dimensional manifolds, within the very high-dimensional space of possible actions, upon which human performances actually tend to lie. This manifold or constraint surface is determined by building two mappings: one from a high-dimensional ``raw-data space'' to a lower-dimensional ``feature space,'' and another from the feature space back to the raw-data space. A best-fit trajectory is arguably the best one-parameter model for a typical human action. A new method is developed to find such a trajectory by fitting a curve to sampled position and velocity performance data. A new spline smoother is derived which can be used within a specially-adapted version of the principal curves algorithm to find a best-fit curve through phase space. The methods investigated in this thesis are evaluated by using them to model a human grasping motion, to recognize hand gestures in a letter-signing application, and to analyze data collected in a robot teleoperation experiment. These experiments show the effectiveness of local nonparametric methods over global parametric methods, and also show that using both position and velocity information results in better models of action trajectories than using position information alone.