A multi-rate predictive control scheme for a robotic eye/head system integrating visual and self-motion cues

In primates, the vestibular-ocular reflex (VOR) is known to stabilize gaze during head perturbations. Also, the internal brain circuits controlling eye movements are found to operate with neural delays much smaller than delays in visual processing pathways (∼2 ms vs. 150 ms). Based on these biological findings, we present a unified multi-rate biomimetic gaze controller integrating VOR mechanisms (self-motion cues) with tracking (pursuit and saccades) for a robotic head with two cameras. The controller uses automatic parametric switching in shared premotor circuits to alternate between two movements types: slow phase (smooth pursuit) relying on visual feedback, and fast phase (blind corrective jumps) producing nystagmus. Similar to the mammalian gaze system, in the absence of a visual target (in dark), the controller generates vestibular nystagmus in response to sinusoidal passive head inputs. During fixation or tracking of a target (slow phase), a head-motion sensor (VOR) detects head rotation direction and drives the cameras in the opposite direction so that gaze in space remains on the visual target. A multi-rate scheme coupled with input prediction is used to allow the internal controller to operate at 600 Hz, while the visual system is limited to a 30 Hz frame rate. Since the saccade circuit is effectively blind, the higher rate controller operation allows increasing saccade bandwidths without ringing to over 30 Hz. Adding slip and memory (PDI control) in the visual feedback overcomes inherent delays in the visual system, increases the tracking response bandwidth to around 5 Hz, and improves steady state tracking gain.