Future service robots will need the ability to work in unpredicted human environments. These robots should be able to learn autonomously without constant supervision in order to adapt to the environment, different users, and changing circumstances. Exploration of unstructured environments requires continuous detection of new objects and learning about them, ideally like a child, through curiosity-driven interactive exploration. Our research work is aimed to design a developmental approach that enables a humanoid robot to perceive its close environment. We take inspiration from human perception in terms of its functionalities and from infant development in terms of the way of learning, and we propose an approach that enables a humanoid robot to explore its environment pro- gressively, like a child through physical actions and social interaction. Following principles of developmental robotics, we focus on incremental, continuous, and autonomous learning that does not require a prior knowledge about the environment or the robot. The perceptual system starts from segmentation of the visual space into proto-objects as units of attention. The appearance of each proto-object is characterized by low-level features based on color and texture that are considered as complementary features. These low-level features are integrated into more complex features and then, into a multi-view model that is learned incrementally and associated with one physical entity. Entities are then classified into three categories : parts of the robot's body, human parts, and manipulable objects. The categorization approach is based on mutual information between the sensory data and proprioception, and also on motion behavior of physical entities. Once the robot is able to categorize entities, it focuses on interactive object exploration. During interaction, the information acquired about an object's appearance is integrated into its model. Thus, interactive learning enhances the knowledge about objects appearances and improves the informativeness of objects models. The implemented active perceptual system is evaluated on an iCub humanoid robot, learning 20 objects through interaction with a human partner and the robot's own actions. Our system is able to recognize objects with 88.5% success and to create coherent representation models that are further improved by interactive learning.
