Abstracting Visual Percepts to Learn Concepts

To efficiently identify properties from its environment is an essential ability of a mobile robot who needs to interact with humans. Successful approaches to provide robots with such ability are based on ad-hoc perceptual representation provided by AI designers. Instead, our goal is to endow autonomous mobile robots (in our experiments a Pioneer 2DX) with a perceptual system that can efficiently adapt itself to ease the learning task required to anchor symbols. Our approach is in the line of meta-learning algorithms that iteratively change representations so as to discover one that is well fitted for the task. The architecture we propose may be seen as a combination of the two widely used approach in feature selection: the Wrapper-model and the Filter-model. Experiments using the PLIC system to identify the presence of Humans and Fire Extinguishers show the interest of such an approach, which dynamically abstracts a well fitted image description depending on the concept to learn.