Learning to Imitate Human Actions through Eigenposes

Programming a humanoid robot to perform an action that takes into account the robot’s complex dynamics is a challenging problem. Traditional approaches typically require highly accurate prior knowledge of the robot’s dynamics and the environment in order to devise complex control algorithms for generating a stable dynamic motion. Training using human motion capture (mocap) data is an intuitive and flexible approach to programming a robot but direct usage of kinematic data from mocap usually results in dynamically unstable motion. Furthermore, optimization using mocap data in the high-dimensional full-body joint-space of a humanoid is typically intractable. In this chapter, we purposes a new model-free approach to tractable imitation-based learning in humanoids by using eigenposes. The proposed framework is depicted in Fig. 1. A motion capture system transforms the Cartesian positions of markers attached to the human body to joint angles based on kinematic relationships between the human and robot bodies. Then, linear PCA is used to create eigenpose data, which are representation of whole-body posture information in a compact low-dimensional subspace. Optimization of wholebody robot dynamics to match human motion is performed in the low dimensional subspace by using eigenposes. In particular, sensory feedback data are recorded from the robot during motion and a causal relationship between eigenpose actions and the expected sensory feedback is learned. This learned sensory-motor mapping allows humanoid motion dynamics to be optimized. An inverse mapping that maps optimized eigenpose data from the low-dimensional subspace back to the original joint space is then used to generate motion on the robot. We present several results