A nonlinear state estimation framework for humanoid robots

This article proposes a novel nonlinear state estimation framework for humanoid robots based on the dynamics of the center of mass (CoM) and dual-loop Kalman filter(DLKF). It effectively fuses the information from inertial measurement unit(IMU), joint encoders, and foot sensitive resistors (FSRs), and provides state estimates for various gait generation algorithms and dynamic balance controllers with CoM or divergence component of motion (DCM) feedback. Compared to the widely used linear models such as the linear inverted pendulum model (LIPM), the nonlinear dynamics of CoM effectively reduce the process error. However, the humanoid robot is a highly complex dynamic system with multiple links and joints, the dynamics of CoM are also a simplification of the whole body dynamics. As a result, it brings non-zero-mean, non-Gaussian, and correlated process error, which the conventional extend Kalman filter (EKF) cannot handle. To this end, the DLKF is adopted to compensate the process error, thus the estimator is robust to the modeling error caused by simplifications. Furthermore, the invariant extended Kalman filter (IEKF) is used for floating base kinematics estimation, and the force/torque (F/T) sensor which is difficult to integrate on smaller or cheaper robots due to the size and cost is not used in our framework. Finally, our nonlinear state estimation framework improves the accuracy of CoM and DCM estimation in a virtual environment simulation using our self-developed Defensor hydraulic humanoid robot.