Autonomously achieving bipedal locomotion skill via hierarchical motion modelling

In this paper, the issue on how a robot autonomously achieves its motion skills is addressed, and an alternative learning approach based on hierarchical motion modelling is proposed. Within the hierarchical model, each robot motion skill is firstly characterized by a family of trajectories that belong to different layers, where inherent constraints among layers will be great helpful in reducing the searching space. Through utilizing the piecewise monotone cubic interpolation method, those trajectories are then parameterized so that a large number of optimization techniques could be applied possibly in model learning. To further debase the learning complexity so that a online learning process can be obtained, a Design of Experiments based Active Learning (DEAL) is employed, which provides an effective exploring strategy with actively selecting samples from hypothesis space by taking advantages from relations among hypotheses in the searching space. To obtain a more robust solution, a random gradient strategy is adopted to adapt or refine the learned output of DEAL. Since the whole online learning process is completed not only under the trial-and-error paradigm, but also without the using of prior dynamic information, the achieving of robot motion skills could be regarded in a completely autonomous style. Experiments are performed on a physical humanoid robot PKU-HR4, and the results illustrate that the proposed approach is effective and promising, which not only speeds up the convergence of the learning process by taking the merits of layered structure and active learning, but also leads to a better locomotion controller since the physical conditions of the involved real robot are taken into account.