Towards an Efficient Contextual Perception for Humanoid Robot: A Selective Attention-based Approach

The intensive computation required for sensing environment and processing sensing information makes the burden of humanoid robot significantly large. Robots are always flooded by perceptual data. Inspired by biological systems, attention mechanism is introduced in this paper to perform an efficient perception system, which will endow humanoid robots with stronger and more efficient power to sense environment. The connotation of attention window(AW) is extended and a more general and abstract definition of AW is given. Its related four kinds of operations are also discussed. The control policies of attention are described which integrate selective perception and distractor inhibition. The distractor inhibition is used to filter large amount of unrelated information, and the consequent perception efficiency promotion is analyzed. Different attention policies are used to control and adjust the perceptual cost and efficiency. The result of experiment shows that the perceptual efficiency has got a big promotion and the effect is well.