Tech Report CS-05-11
Extensible Data-driven Classification of Robot Sensor Data
Daniel H Grollman and Odest Chadwicke Jenkins and Frank Wood
We present an unsupervised method for online classification of robot sensor data. Unlike previous work in which human intuitive correspondences are sought between sensor data and classes of physical space (room, wall, corner, door, etc.), our method learns intrinsic classes based on the characteristics of sensor data. We approximate the manifold underlying sensor data using Isomap nonlinear dimension reduction and use classical Bayesian clustering (Gaussian mixture models) with model identification techniques to discover classes. The learned model can then be used to classify new (out-ofsample) data in real time. We apply our method to sensor data of different modalities and from different physical spaces. Our results demonstrate the robustness of the method with respect to noise and robot location variance. The extensibility of our approach allows us to classify large and streaming data sets. We conclude that data-driven classification may serve as a more solid foundation for applications that require the ability to discriminate physical spaces surrounding a robot.
(complete text in pdf)