A robotic camera platform for evaluation of biomimetic gaze stabilization using adaptive cerebellar feedback

Abstract: This thesis describes the development of a robotic platform for evaluation of gaze stabilization algorithms built for the Sensorimotor Systems Laboratory at the University of British Columbia. The primary focus of the work was to measure the performance of a biomimetic vestibulo-ocular reflex controller for gaze stabilization using cerebellar feedback. A flexible robotic system was designed and built in order to run reproducible test sequences at high speeds featuring three dimensional linear movement and rotation around the vertical axis. On top of the robot head a 1 DOF camera head can be independently controlled by a stabilization algorithm implemented in Simulink. Vestibular input is provided by a 3-axis accelerometer and a 3-axis gyroscope. The video feed from the camera head is fed into a workstation computer running a custom image processing program which evaluates both the absolute and relative movement of the images in the sequence. The absolute angles of tracked regions in the image are continuously returned, as well as the movement of the image sequence across the sensor in full 3 DOF camera rotation. Due to dynamic downsampling and noise suppression algorithms very good performance was reached, enabling retinal slip estimation at 720 degrees per second. Two different controllers were implemented, one adaptive open loop controller similar to Dean et al.’s work[12] and one reference implementation using closed loop control and optimal linear estimation of reference angles. A sequence of tests were run in order to evaluate the performance of the two algorithms. The adaptive controller was shown to offer superior performance, dramatically reducing the movement of the image for all test sequences, while also offering better performance as it was tuned over time.