Abstract

We derive categories directly from robot sensor data to address the symbol grounding problem. Unlike model-based approaches where human intuitive correspondences are sought between sensor readings and facets of an environment (corners, doors, etc.), our method learns intrinsic categories (or natural kinds) from the raw data itself. We approximate a manifold underlying sensor data using Isomap nonlinear dimension reduction and use Bayesian clustering (Gaussian mixture models) with model identification techniques to discover kinds. Applying our technique to sensor data of different modalities and from different physical spaces we demonstrate robustness with respect to noise and robot location. We also demonstrate a method for applying learned kinds to new sensor data (out-of-sample readings) in real time to show the efficacy of our technique as a foundation for topological mapping and autonomous control. Lastly, we discuss the application of our technique toward massive (250,000 datapoint) data sets. 1.

Citations