Learning dynamic humanoid motion using predictive control in low dimensional subspaces

Imitation of complex human motion by a humanoid robot has long been recognized as an important problem in robotics. The problem is particularly difficult when body dynamics such as balance and stability must be taken into account during imitation. In this paper we present a framework applicable to the problem of imitating an input motion while simultaneously considering dynamic motion stability. Our framework leverages two main components. Firstly, dimensionality reduction techniques allow for efficient and compact state and control signal representations. Secondly, a learning-based predictive control architecture generates novel motions optimizing over expected sensory signals. We demonstrate results on modifying an input walking gait which allows for both faster and more stable walking