Autonomous Learning Evaluation toward Active Motor Babbling

Learning in robotics is one of the practical solutions allowing an autonomous robot to perceive its body and the environment. As discussed in the context of the frame problem [1], the robot’s body and the environment are too complex to be modeled. Even if the kinematics and the dynamics of the body are known, a real sensory input to the body would be different to one derived from the theoretical model, because the sensory input is always influenced by the interaction with the environment. For instance, when we grasp an object, the physical state of our arm such as a weight and momentum becomes different to those at the normal state. However, it is difficult to evaluate all potential variation in advance, since real data can vary quite a lot and the behavior of the external environment is not necessarily controlled by the robot: in this example, the state of the arm is always different depending on the grasped object. On the other hand, learning provides a data-driven solution: the robot explores the environment and extracts knowledge to build an internal model of the body and the environment. Learning-based motor control system is well studied in the literature [2] [3] [4] [5] [6] [7]. Haruno et al. proposed a modular control approach [3], which couples a forward model (state predictor) and an inverse model (controller). The forward model predicts the next state from a current state and a motor command (an efference copy), while the inverse model generates a motor command from the current state and the predicted state. The desired motor command is not available, but the feedback error learning procedure (FEL) provides a suitable approximation [4]. The prediction error contributes to gate learning of the forward and inverse models, and to weight output of the inverse models for the final motor command. Motor prediction based on a copy of motor command compensates the delays and noise in the sensorimotor system. Moreover, motor prediction allows differentiating self-generated movements from externally imposed forces/disturbances [5][6]. Learning-based perception is applicable not only for motor control but also to model the environment owing to multiple sensorial modalities, such as vision, audition, touch, force/torque, and acceleration sensing. In a similar approach, we developed a learning system aiming at predicting future sensing data based