Using eigenposes for lossless periodic human motion imitation

Programming a humanoid robot to perform an action that takes the robot's complex dynamics into account is a challenging problem. Traditional approaches typically require highly accurate prior knowledge of the robot's dynamics and environment in order to devise complex control algorithms for generating a stable dynamic motion. Training using human motion capture is an intuitive and flexible approach to programming a robot but directly applying motion capture data to a robot usually results in dynamically unstable motion. Optimization using high-dimensional motion capture data in the humanoid full-body joint-space is also typically intractable. In previous work, we proposed an approach that uses dimensionality reduction to achieve tractable imitation-based learning in humanoids without the need for a physics-based dynamics model. This work was based on a 3-D “eigenpose” representation. However, for some motion patterns, using only three dimensions for eigenposes is insufficient. In this paper, we propose a new method for motion optimization based on high-dimensional eigenpose data. A one-dimensional computationally efficient motion-phase optimization method is implemented along with a newly developed cylindrical coordinate transformation technique for hyperdimensional subspaces. This results in a fast learning algorithm and very accurate motion imitation. We demonstrate the new algorithm on a Fujitsu HOAP-2 humanoid robot model in a dynamic simulator and show that a dynamically stable sidestep motion can be successfully learned by imitating a human demonstrator.