Observing human movements to construct a humanoid interface

We introduce a novel approach to learn a humanoid interface by using observed human behaviors. We propose using the observed human movements to extract task-relevant degrees-of-freedoms (DOFs) so that we can construct the humanoid interface to generate high-dimensional humanoid movements by using low-dimensional user inputs. The extracted intrinsic DOFs are represented as a task-relevant manifold. On the other hand, since the manifold is derived from the observed human movements, we cannot directly use the movements generated through the manifold to control a humanoid. Therefore, we introduce a calibration procedure to convert the movements generated through the manifold to the humanoid robot movements. By using the task-relevant manifold and the movement conversion, we can control a many-DOF humanoid robot using a low-dimensional command input interface such as a game pad. We show that fourteen-degrees-of-freedom humanoid robot can be controlled to draw two-dimensional spiral and star shapes in the three-dimensional Cartesian space by using the proposed low-dimensional control interface.