Predictive Whole-Body Control of Humanoid Robot Locomotion

Humanoid robots are machines built with an anthropomorphic shape. Despite decades of research into the subject, it is still challenging to tackle the robot locomotion problem from an algorithmic point of view. For example, these machines cannot achieve a constant forward body movement without exploiting contacts with the environment. The reactive forces resulting from the contacts are subject to strong limitations, complicating the design of control laws. As a consequence, the generation of humanoid motions requires to exploit fully the mathematical model of the robot in contact with the environment or to resort to approximations of it. This thesis investigates predictive and optimal control techniques for tackling humanoid robot motion tasks. They generate control input values from the system model and objectives, often transposed as cost function to minimize. In particular, this thesis tackles several aspects of the humanoid robot locomotion problem in a crescendo of complexity. First, we consider the single step push recovery problem. Namely, we aim at maintaining the upright posture with a single step after a strong external disturbance. Second, we generate and stabilize walking motions. In addition, we adopt predictive techniques to perform more dynamic motions, like large step-ups. The above-mentioned applications make use of different simplifications or assumptions to facilitate the tractability of the corresponding motion tasks. Moreover, they consider first the foot placements and only afterward how to maintain balance. We attempt to remove all these simplifications. [continued]