Learning Capture Points for Humanoid Push Recovery

We present a method for learning Capture Points for humanoid push recovery. A Capture Point is a point on the ground to which the biped can step and stop without requiring another step. Being able to predict the location of such points is very useful for recovery from significant disturbances, such as after being pushed. While dynamic models can be used to compute Capture Points, model assumptions and modeling errors can lead to stepping in the wrong place, which can result in large velocity errors after stepping. We present a method for computing Capture Points by learning offsets to the Capture Points predicted by the Linear Inverted Pendulum Model, which assumes a point mass biped with constant Center of Mass height. We validate our method on a three dimensional humanoid robot simulation with 12 actuated lower body degrees of freedom, distributed mass, and articulated limbs. Using our learning approach, robustness to pushes is significantly improved as compared to using the Linear Inverted Pendulum Model without learning.