CAIAC

This thesis develops novel computational models and robots to investigate how move- ments are controlled by the human brain. The thesis focuses on the control of eye movements as the eye is a relatively simple, yet complete, sensory and motor system in the human body. In particular, we look at two types of eye movements: gaze shifting and gaze stabilization.

The eye constantly shifts gaze by performing very fast, yet accurate, movements. These movements (called saccades) are too fast to control using the usual engineering approach of relying on feedback. This stems from the delays inherent in our nervous system; a saccade of less than 10 degrees can be completed in less than 50 milliseconds, while low level visual processing alone takes 40 milliseconds. To achieve fast movements in the face of large delays, we develop a controller that learns a non-linear dynamic model of the eye. It then uses the "pulse-step" model proposed in neuroscience to compute the activation signal that drives the saccade. We tested the controller on a robotic eye we developed, with complicated non-linearities, and obtain performance comparable to the human eye. Interestingly, it also reproduces other characteristic relationships observed in human eye movements.

Another type of eye movement investigated in the thesis is gaze stabilization, which plays an important role in enabling clear vision as we move about in the environment. It com- pensates for head movement by incorporating information about head movement measured by the fast vestibular system in our inner ears and the slower visual information from the eye. We developed a controller that learns a model for generating the appropriate activation signals for stabilizing gaze and implemented it on a robotic head platform mounted on a high speed robotic body. The "head" contains the equivalent of an eye and vestibular system. We demonstrate the controller's effectiveness in stabilizing the camera in the face of high speed perturbations, and demonstrate the benefits for computer vision algorithms on fast mobile robots.