We have been studying the control of eye movements for the
Babybot. Various behaviors were implemented that allow the robot to
visually explore the environment in search of interesting objects and
events. The simplified Babybot's visual system can detect color and
motion. The robot also uses binocular disparity to gather information
about the depth in the visual scene and control vergence.
The
head is embedded with three gyroscopes that are used by the Babybot to
develop the sense of a stable world. Visual stability of the world is of
course important to perceive it correctly: the visual processing is
simpler if the eyes are not moving too much. Also, these signals can be
used to actively coordinate the movement of the head with that of the eyes
(compensatory eye movements). Experiments investigated different types of
self-supervised learning to automatically tune the performance of this
class of eye movements.
We carried out experiments on
learning saccade (fast eye movements) towards visually identified targets.
Neural networks techniques were employed for this task.
Controlling
the movement of the eyes and head is one of the first step required in
learning complex cognitive tasks. It also rests at the foundation of
reaching and manipulation.