Unsupervised identification and recognition of situations for high-dimensional sensori-motor streams

An important question in self-learning robots is how robots can autonomously learn about and act in their environment in an on-line and unsupervised manner. This paper introduces and evaluates Context Recognition in Data Streams (CoRDS), a method that enables a robot to identify and recognise different situations in its environment. CoRDS achieves this by processing the data stream from the robot's sensors to distinguish different patterns that identify different environmental situations. We evaluated the CoRDS method by means of quantitative and qualitative analysis on three different data streams: one synthetic and two data sets with actual sensor data generated by Thymio II robots. Our analyses showed that CoRDS created active cluster patterns that, for all three data streams, corresponded with the experimenters' expectations. Experiments varying the parameters of the CoRDS method indicated a consistent response over all three data streams. These findings suggest that CoRDS may provide a basis for data stream clustering techniques that can be applied for the task of situation recognition.