State Estimation for Humanoid Robots

Abstract

This thesis focuses on dynamic model based state estimation for hydraulic humanoid robots. The goal is to produce state estimates that are robust and achieve good performance when combined with the controller. Three issues are addressed in this thesis. How to use force sensor and IMU information in state estimation? How to use the full-body dynamics to estimate generalized velocity? How to use state estimation to handle modeling error and detect humanoid falling? Hydraulic humanoid robots are force-controlled. It is natural for a controller to produce force commands to the robot using inverse dynamics. Model based control and state estimation relies on the accuracy of the model. We address the issue: “To what complexity do we have to model the dynamics of the robot for state estimation?”. We discuss the impact of modeling error on the robustness of the state estimators, and introduce a state estimator based on a simple dynamics model, it is used in the DARPA Robotics Challenge Finals for fall detection and prevention. Hydraulic humanoids usually have force sensors on the joints and end effectors, but not joint velocity sensors because there is no high velocity portion of the transmission as there are no gears. A simple approach to estimate joint velocity is to differentiate measured joint position over time and low pass filter the signal to remove noise, but it is difficult to balance between the signal to noise ratio and delay. To address this issue, we will discuss three ways to use the full-body dynamics model and force sensor information to estimate joint velocities. The first method efficiently estimates the full state through decoupling. It estimates the base variables by fusing inertial sensing with forward kinematics, and joint variables using forward dynamics. The second method estimates the generalized velocity using quadratic program. Force sensor information is also taken into account as an optimization variable in this formulation. The third method uses low cost MEMS IMUs to measure link angular velocities, and integrate that information into joint velocity estimation. Some of these state estimators were used on the Atlas robot for full body control, odometry and fall detection and prevention. In the DARPA Robotics Challenge Finals, we achieved 12/14 points and had no fall or human intervention.