Humanoid Standing Control: Learning from Human Demonstration

A three-dimensional numerical model of human standing is presented that reproduces the dynamics of simple swaying motions while in double-support. The human model is structurally realistic, having both trunk and two legs with segment lengths and mass distributions defined using human morphological data from the literature. In this investigation, model stability in standing is achieved through the application of a high-level, reduced-order control system where stabilizing forces are applied to the model’s trunk by virtual springdamper elements. To achieve biologically realistic model dynamics, torso position and ground reaction force data measured on human subjects are used as demonstration data in a supervised learning strategy. Using Powell’s method, the error between simulation data and measured human data is minimized by varying the virtual high-level force field. Once optimized, the model is shown to track torso position and ground reaction force data from human demonstrations. With only these limited demonstration data, the humanoid model sways in a biologically realistic manner. The model also reproduces the center-of-pressure trajectory beneath the foot, even though no error term for this is included in the optimization algorithm. This indicates that the error terms used (the ones for torso position and ground reaction force) are sufficient to compute the correct joint torques such that independent metrics, like center-of-pressure trajectory, are correct.