Motor biases in visual attention for a humanoid robot

摘要

Tantalizing evidence derived from psychophysics and developmental psychology experiments has shown that attention is task-dependent. Two characteristics of human control of attention are very relevant for humanoid robots, namely, the ability to predict the context (task dependence) from the observed stimuli, and the ability to learn an appropriate movement strategy perhaps over developmental time scales. In this paper we aim at implementing these features to control attention in a humanoid robot by including a set of trajectory predictors in the simple but effective form of Kaiman filters, and, more importantly, a reinforcement learning based process that utilizes the predictors and the complete set of actions of the robot repertoire to generate a suitably optimal action sequence. Preliminary experiments show that the system indeed works correctly: it uses the predictors to discriminate the environmental context (e.g. static vs. dynamic) and produces a valid control policy that drives the robot to fixation of the task-relevant static or moving stimuli.