Spectral Subtraction of Robot Motion Noise for Improved Event Detection in Tactile Acceleration Signals

New robots for teleoperation and autonomous manipulation are increasingly being equipped with high-bandwidth accelerometers for measuring the transient vibrational cues that occur during contact with objects. Unfortunately, the robot's own internal mechanisms often generate significant high-frequency accelerations, which we term ego-vibrations. This paper presents an approach to characterizing and removing these signals from acceleration measurements. We adapt the audio processing technique of spectral subtraction over short time windows to remove the noise that is estimated to occur at the robot's present joint velocities. Implementation for the wrist roll and gripper joints on a Willow Garage PR2 robot demonstrates that spectral subtraction significantly increases signal-to-noise ratio, which should improve vibrotactile event detection in both teleoperation and autonomous robotics.