Active Motor Babbling for Sensory-Motor Learning
Résumé
For a complex autonomous robotic system such as a humanoid robot, the motor-babbling based sensory-motor learning is considered effective to develop an internal model interaction between the body and the environment interaction. In this paper we investigate a method of learning evaluation which directs the exploration to the learning domain of interest. The model is characterized by a function called confidence, which returns a high value if the measured sensory state and motor command match the predicted state and the estimated motor command for state control. The learning system creates maps of the confidence for states regarding the state prediction and the state transition, which allows the robot to notice if the current sensory state is predictable and controllable, or not. We consider the confidence function as the first step to the behavior design of the robot expecting autonomous adaptation for the environment. The approach was experimentally validated using the humanoid robot James.