IMU-based joint state estimation for humanoid control

Feedback control for robots relies on accurate estimates of the joint state. Traditionally, the joint position is measured using an angular sensor from which joint derivatives are computed numerically. This amplifies noise, requiring low-pass filtering for use in control. However, filtering introduces a delay, preventing the use of feedback gains high enough for satisfactory stiffness and damping. We have developed methods [1] for computing joint velocity and acceleration from inertial sensors which measure quantities of the same order. These estimates-from gyroscopes and accelerometers-can be used for higherperformance robot control and potentially even for determining the joint state of a human subject.