The information used to determine the internal model of a robot system emerges from individual interactions with the environment. Knowledge about a specific internal model can be acquired by means of model learning techniques based on supervised machine learning tools, such as neural networks. One of the main challenges is to specify the starting point for the learning process. We propose a methodology for initial weight estimation based on the robot morphology. We use the generation of saccades in a robotic head as a case study to evaluate the performance of this approach. Our results suggest that this methodology, based on learning the relationship between the morphological parameters of the robot and those in its internal model, improves the quality of weight initialization, resulting in a considerable speed-up in the process of learning the internal model.

Saccades are fast eye movements that allow humans and robots to bring the visual target in the center of the visual field. Saccades are open loop with respect to the vision system, thus their execution require a precise knowledge of the internal model of the oculomotor system. In this work, we modeled the saccade control, taking inspiration from the recurrent loops between the cerebellum and the brainstem. In this model, the brainstem acts as a fixed-inverse model of the oculomotor system, while the cerebellum acts as an adaptive element that learns the internal model of the oculomotor system. The adaptive filter is implemented using a state-of-the-art neural network, called I-SSGPR. The proposed approach, namely recurrent architecture, was validated through experiments performed both in simulation and on an antropomorphic robotic head. Moreover, we compared the recurrent architecture with another model of the cerebellum, the feedback error learning. Achieved results show that the recurrent architecture outperforms the feedback error learning in terms of accuracy and insensitivity to the choice of the feedback controller.

Active eye movements can be exploited to learn the oculomotor command that allows a humanoid head to gaze at a visual target. In the field of adaptive robotics this transformation is commonly encoded by means of neural networks that are trained using ballistic eye movements. In this work, we encode the visuo-oculomotor transformation using radial basis function networks and we derive the equations to use such a transformation in visual servo control tasks. We study how the choice of the input (monocular vs binocular) influences the execution of ballistic and visually-guided eye movements. Analytic analysis and experimental results on a simulated model of the robot show that the proposed framework can be employed to execute both kind of movements.