WHAT DO OBJECTS FEEL LIKE? Active Perception for a Humanoid Robot

We present a recurrent neural architecture with parametric bias for actively perceiving objects. A humanoid robot learns to extract sensorimotor laws and based on those to classify eight objects by exploring their multi-modal sensory characteristics. The network is either trained with prototype sequences for all objects or just two objects. In both cases the network is able to self-organize the parametric bias space into clusters representing individual objects and due to that, discriminates all eight categories with a very low error rate. We show that the network is able to retrieve stored sensory sequences with a high accuracy. Furthermore, trained with only two objects it is still able to generate fairly accurate sensory predictions for unseen objects. In addition, the approach proves to be very robust against noise.