Robust State Estimation With Redundant Proprioceptive Sensors

We present a framework for robust estimation of the configuration of an articulated robot using a large number of redundant proprioceptive sensors (encoders, gyros, accelerometers) distributed throughout the robot. Our method uses an Unscented Kalman Filter (UKF) to fuse the robot’s sensor measurements. The filter estimates the angle of each joint of the robot, enabling the accurate estimation of the robot’s kinematics even if not all modules report sensor readings. Additionally, a novel outlier detection method allows the the filter to be robust to corrupted accelerometer and gyro data.