Learning a humanoid robot interface by embedding a low-dimensional command manifold into a high-dimensional joint action space

In this paper, we propose a novel way of constructing a humanoid robot interface by embedding a low-dimensional command space into observed high-dimensional joint angle action space. It is almost impossible for users to independently and simultaneously control all the joints of a humanoid robot. On the other hand, for a given target task, not all the degrees of freedom (DOFs) of the robot may necessarily be used. The task can be accomplished by using fewer properly selected DOFs. In our approach, we embed the low-dimensional command manifold into the original high-dimensional joint angle space. For the command manifold embedding, we use Locally Smooth Manifold Learning (LSML) by which we can find high-dimensional tangent vectors on the low-dimensional command manifold. By using the derived tangent space, we can embed the low-dimensional control command that can be specified by a few-DOF gamepad into joint angle movements of the humanoid robot. We show that both simulated and a real 14-DOF humanoid robots can be efficiently controlled by using a 2-DOF gamepad with our proposed interface.