Learning of Action Generation from Raw Camera Images in a Real-World-Like Environment by Simple Coupling of Reinforcement Learning and a Neural Network

摘要

For the development of human-like intelligent robots, we have asserted the significance to introduce a general and autonomous learning system in which one neural network simply connects from sensors to actuators, and which is trained by reinforcement learning. However, it has not been believed yet that such a simple learning system actually works in the real world. In this paper, we show that without giving any prior knowledge about image processing or task, a robot could learn to approach and kiss another robot appropriately from the inputs of 6240 color visual signals in a real-world-like environment where light conditions, backgrounds, and the orientations of and distances to the target robot varied. Hidden representations that seem useful to detect the target were found. We position this work as the first step towards taking applications of the simple learning system away from "toy problems".