Autonomous learning based on depth perception and behavior generation

We propose a new neuro-robotic network that can simultaneously achieve a goal oriented behavior task and perception enhancement task for a visually-guided object manipulation based on learning by examples. The brain exploits action to develop perception qualities, and perceptual process helps to develop qualified-behavior. In order to import those action and perception inter-abilities of a brain into a humanoid robot, we consider two key inspirations: (1) Sensory Invariant Driven Action (SIDA) and (2) Object Size Invariance (OSI) characteristic. Considering robot manipulation of a target object with distance estimation as a perceptual process, we develop a new autonomous learning method based on the SIDA for behavior generation and OSI property for perceptual judgment. The proposed method is evaluated by using a humanoid robot (NAO) with stereo cameras, and the experimental results show that the proposed method is effective on autonomously improving the behavior generation performance as well as depth perception accuracy.