Sensory Prediction Learning - How to Model the Self and the Environment

摘要

For a complex autonomous robotic system such as a humanoid robot, the learning-based state prediction is considered effective to develop the body and environment model autonomously. In this paper we investigate a model of changes detection directly included in the evaluation process of the learning algorithm. The model is characterized by a function called confidence, which returns a high value if the robot's actual state data match the predicted state data. The robot then creates the confidence map for each sensor based on the prediction error, which allows the robot to notice if the current sensory state is predictable (experienced) or not. We consider the confidence function as the first step to self diagnosis and self adaptation. The approach was experimentally validated using the humanoid robot James.