Vision-based motion primitives for reactive walking

This paper presents a method for reactive walking allowing visual servoing and adaptation of footsteps trajectories in real-time. This is done by building upon recent advances in the fields of optimal control for a walking pattern generator [1] and planning for a nonholomic robot with field-of-view constraints [2]. Herdt et al. [1] provided a controller where a humanoid robot is driven by its Center-of-Mass (CoM) velocity. A natural extension, proposed in [3], is to control a humanoid robot directly by using vision-based control techniques. However, the trajectories realized by the robot in that case are generated to minimize the distance in the image feature space and might create unnecessary motion in the space of the footprints. This paper aims at solving this problem by making the CoM follow a convenient space of trajectories for which the robot behaves overall better. The motion primitives obtained in [2] provide the space of trajectories used in this work.