Humanoid Robot Balance Control

Studies of human standing balance have identified multiple discrete strategies: the 'ankle strategy', in which all joints except for the ankle are fixed and torque about the ankle joint is used to compensate for the perturbation, The ' hip strategy', in which torque about the hip joint is used to accelerate the torso and move the Center of Mass (CoM), the 'arm strategy' and the 'step strategy' when the perturbation is very large. For a constant push, humans will lean to the push and use the gravity to compensate for it.

Standing balance control is an important control problem for humanoid robots. It is shown that a single optimization criterion can be used to generate multiple balance recovery strategies. We employ a library of optimal trajectories and the neighboring optimal control method to compute local approximations to the optimal control. We take advantage of a parametric nonlinear optimization method, SNOPT, to generate initial trajectories and then use Differential Dynamic Programming (DDP) to refine them. A library generation method is proposed, which keeps the trajectory library to a reasonable size. We compare the proposed controller with an LQR based gain scheduled controller with the same optimization criterion. Simulation and experiment results demonstrate the performance of the proposed method.