Learn Like Infants: A Strategy for Developmental Learning of Symbolic Skills Using Humanoid Robots

To interact with the operator intuitively, a robot must build symbolic representations of the environment. For this purpose, unsupervised methods are insufficient in modeling semantic information, and supervised methods are inefficient for general applications. To solve the problem, we develop an incremental learning strategy by imitating the learning process of human infants, described by developmental psychology theory. This theory divides the infant learning process into four stages, from initial sensorimotor learning to high level intelligence. Inspired by these stages, we describe the developmental robotic object learning with two consecutive processes, composed of sample-based learning and symbolic learning. In the first process, the robot manipulates the target objects to build sample-based representations, and uses particle filter to update the object models after sequential manipulations. With the sample-based object representations, the robot uses latent support vector machine to learn part-based object models, thus it can recognize the objects accurately and interact with the operator intuitively in practical tasks. We implement our strategy with a humanoid robot, and demonstrate its incremental learning of symbolic representations of rigid objects and articulated objects. The result shows that our method allows the robot to symbolically represent various objects more autonomously, and to recognize reappearing objects for interaction with improving accuracy.