Sensory-objects network driven by intrinsic motivation for survival abilities

An autonomous robotic system must be able to learn the meaning of objects in its surrounding in terms of both survivability and task accomplishment. In this paper, a new type of neural network in which objects of the environment, physiological parameters as well as possible actions are represented by neurons, is proposed as a robot controller. This approach is evaluated by controlling a simulated robot playing a game in which it has to maximize its score while harvesting the energy necessary for its survival. When the game starts, the robot does not know what action on what object will bring points or energy, making learning abilities necessary to achieve a high score. Also the game has been designed such that objects in the environment are characterized by a continuous property that can not be easily categorized. The environment is also dynamic in the sense that an object´s reserve in points or energy can get depleted.