Active Motor Babbling for Sensory-Motor LearningRyo Saegusa, Giorgio Metta, and Giulio SandiniRobotics, Brain and Cognitive Sciences DepartmentItalian Institute of TechnologyVia Morego 30, 16163 Genoa, Italyryosieee.org, pasaliralab.it, giulio.sandiniiit.itSophie SakkaLaboratory of Solids MechanicsUniversity of PoitiersBP30179-86962 Futuroscope Chasseneuil Cedex, Francesophie.sakkalms.univ-poitiers.frAbstractFor a complex autonomous robotic system suchas a humanoid robot, the motor-babbling based sensory-motorlearning is considered effective to develop an internal modelof the self-body and the environment autonomously. In thispaper we propose a methodology of sensory-motor learningand its evaluation towards active learning. The proposed modelis characterized by a function called confidence, which worksas a memory of reliability for state prediction and control.The confidence for the state can be a good measure to biasthe next exploration strategy of data sampling, such as todirect its state to the unreliable domain. We consider theconfidence function as the first step to an active behavior designfor autonomous environment adaptation. The approach wasexperimentally validated using the humanoid robot James.Index TermsSensory motor prediction, Neural networks,Learning, humanoid robot, ConfidenceI. INTRODUCTIONLearning in robotics is one of the practical solutionsallowing an autonomous robot to perceive its body andthe environment. As discussed in the context of the frameproblem 1, the robots body and the environment are toocomplex to be modeled. Even if the kinematics and thedynamics of the body are known, a real sensory input to thebody would be different to one derived from the theoreticalmodel, because the sensory input is always influenced bythe interaction with the environment. For instance, whenwe grasp an object, the physical state of our arm such asa weight and momentum becomes different to those at thenormal state. However, it is difficult to evaluate all potentialvariations in advance, since real data can vary quite a lot andthe behavior of the external environment is not necessarilycontrolled by the robot: in this example, the state of the armis always different depending on the grasped object. On theother hand, learning provides a data-driven solution: the robotexplores the environment and extracts knowledge to build aninternal model of the body and the environment.Learning-based motor control systems are well studied inthe literature 2 3 4 5 6 7. Haruno et al. proposedThe work presented in this paper was partially supported by theROBOTCUB project (IST-2004-004370) and the CONTACT project (NEST-5010), funded by the European Commission through the Unit E5 CognitiveSystems.a modular control approach 3, which couples a forwardmodel (state predictor) and an inverse model (controller).The forward model predicts the next state from a currentstate and a motor command (an efference copy), whilethe inverse model generates a motor command from thecurrent state and the predicted state. Even if a proper motorcommand is unknown, the feedback error learning procedure(FEL) provides a suitable approximation 4. The predictionerror contributes to gate learning of the forward and inversemodels, and to weight output of the inverse models for thefinal motor command. Motor prediction, based on a copyof the motor command, compensates the delays and noise inthe sensory-motor system. Moreover, motor prediction allowsdifferentiating self-generated movements from externally im-posed forces/disturbances 56.Learning-based perception is applicable not only for motorcontrol but also to model the environment owing to mul-tiple sensorial modalities, such as vision, audition, touch,force/torque, and acceleration sensing. In a similar approach,we developed a learning system aiming at predicting futuresensing data based on current sensing data and motor com-mand 8. Unlike most studies on sensory-motor prediction,the robot and the environment are considered dynamic. Thus,we explored the possibilities for the robot to detect changesin its self or its environment in an autonomous manner:no other information such as a model was given to thesystem. Following this concept, we investigated a functioncalled confidence, driven in the evaluation process of thesensory prediction learning 9. The aim of this function isto detect inequalities between the predicted situation and thereal situation of the body and the environment. The notion ofrobotic self-confidence was developed as the first step towardself diagnosis and self adaptation.Our global aim is to implement a learning process as anatural adaptation and self-improvement for the robot. In thiscontext, one of the significant problems in learning is that itrequires much time for data sampling and post treatment. Anefficient learning strategy is necessary to enhance the learningspeed while keeping its quality. The random sampling strat-egy is considered as th