Learning-based methods for planning and control of humanoid robots

Nowadays, humans and robots are more and more likely to coexist as time goes by. The anthropomorphic nature of humanoid robots facilitates physical human-robot interaction, and makes social human-robot interaction more natural. Moreover, it makes humanoids ideal candidates for many applications related to tasks and environments designed for humans. No matter the application, an ubiquitous requirement for the humanoid is to possess proper locomotion skills. Despite long-lasting research, humanoid locomotion is still far from being a trivial task. A common approach to address humanoid locomotion consists in decomposing its complexity by means of a model-based hierarchical control architecture. To cope with computational constraints, simplified models for the humanoid are employed in some of the architectural layers. At the same time, the redundancy of the humanoid with respect to the locomotion task as well as the closeness of such a task to human locomotion suggest a data-driven approach to learn it directly from experience. This thesis investigates the application of learning-based techniques to planning and control of humanoid locomotion. In particular, both deep reinforcement learning and deep supervised learning are considered to address humanoid locomotion tasks in a crescendo of complexity. First, we employ deep reinforcement learning to study the spontaneous emergence of balancing and push recovery strategies for the humanoid, which represent essential prerequisites for more complex locomotion tasks. Then, by making use of motion capture data collected from human subjects, we employ deep supervised learning to shape the robot walking trajectories towards an improved human-likeness. The proposed approaches are validated on real and simulated humanoid robots. Specifically, on two versions of the iCub humanoid: iCub v2.7 and iCub v3.