UC San Diego

Biologically inspired humanoid robots present new challenges for system identification and control due to the presence of many degrees of freedom, highly compliant actuators, and non-traditional force transmission mechanisms. In this thesis, we address these challenges using machine learning approaches. The key idea is to replace classical laborious manual model calibration and motion programming with statistical inference and learning from multi-modal sensory data. To this end, we develop several new parametric models and their parameter identification algorithms enabling new sensor/ actuator configurations beyond the scope of previous approaches. In addition, we also develop a semi-parametric model to learn from experiences not predicted by the parametric model. Using similar approaches grounded in machine learning, we also develop methods to allow humanoid robots to learn to make facial expressions, kick a ball, and to reach for objects while collaborating with people. We collected a unique dataset that describes development of infant reaching behavior while interacting with an adult caregiver. We compared the observed development of social reaching in human infants with the machine learning based development behavior in a complex humanoid robot