Dynamic humanoid balance through inertial control

Physical humanoids often require the ability to maintain upright balance while performing various tasks involving locomotion and environmental interaction. Such balance requirements have been difficult to maintain with traditional approaches to articulated motion control. We claim that these difficulties are significantly due the use of parent-space in standard proportional-derivative (PD) servoing, typically requiring highly sophisticated decision making policies to function while maintaining balance. Using inspiration from inverted pendulum robots, we address humanoid balance control through a world-space servoing model. Our model retains the same basic form as the PD-servo, but uses inertial/accelerometer measurements rather potentiometer-like sensing. Our humanoids are able to functionally balance, locomote, and recover without sophisticated decision making. We demonstrate the efficacy of our approach through simulation experiments involving locomotion, user interaction, ballistic motion, uneven terrain, and dramatic disturbances.