Learning Periodic Human Motion

Through Imitation Using Eigenposes

 

This work provides the first demonstration that a humanoid robot can learn to perform human dynamic motion such as walking directly by imitating a human gait obtained from motion capture data without any prior information of its dynamics model. 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 environment in order to devise complex (and often brittle) control algorithms for generating a stable dynamic motion.  Training using human motion capture is an intuitive and flexible approach to programming a robot but direct usage of mocap data usually results in dynamically unstable motion. Furthermore, optimization using mocap  data in the humanoid full-body joint-space is typically intractable. This work purposes a new model-free approach to tractable imitation-based learning in humanoids. Kinematic information from human motion capture is represented in a low dimensional subspace. We call this compact representation of posture data “Eigenposes”. Eigenposes are mapped with sensory feedback to form a predictive dynamic model. This model is used within an optimization framework to estimate optimal motor commands that satisfy the initial kinematic constraints as best as possible while at the same time generating dynamically stable motion. The viability of this approach is demonstrated by providing examples of dynamically stable walking learned from motion capture data using both a simulator and a real humanoid robot.

 

The watermill examples

Complicate humanoid motion can be governed by very few parameters. For the cases in this video, there is only one single parameter that drives the motion which is the mill.

Eigenpose

A simple low-dimensional representation of human motion can be created by using linear components analysis (PCA). We call these low-dimensional posture data “Eigenposes”.

The action subspace embedding

A closed-curve 3D eigenpose data pattern can be constrained in a cylindrical  coordinate system as a single parameter function:

Optimized eigenposes

Causal relationship of eigenposes and sensory feedback can be discover by learning algorithm to construct dynamics model for stable motion optimization.

Learning to walk through imitation

3D eigenposes and gyroscope feedback were used for HOAP-2 humanoid robot to learn walking motion from motion capture data.

Lossless motion imitation

An extension concept of coordinate transformation between Cratesian coordinate system and cylindrical coordinate system for hyper-dimensional cases let us achieve lossless motion optimization through single parameter function of eigenposes.