Multimodal Sparse Representation for Anomaly Classification in A Robot Introspection System

Robot introspection is expected to greatly aid longer-term autonomy of autonomous systems. By equipping robots with skills that allow them to assess the quality of their sensory data, robots can detect and classify anomalies and recover from them. However, the variability and complexity of multimodal sensory data exacerbates anomaly classification in practical applications. This work contributes a sparse representation for multimodal time-series for relevant features while preserving classification accuracy. We examine multivariate features in the time and frequency domains and analyze the correlation across all sensory signal dimensions. The sparse process consists of four phases: (i) preprocessing and synchronization of raw time signals; (ii) statistical formulations for each sensor, (iii) evaluation of feature significance using hypothesis testing, and (iv) raw multimodal sample representation with the concatenating feature vector. We test the performance of the sparse anomaly classification through the enactment of a kitting experiment designed to package 6 different objects from a given placement region to a specific container. Aaccuracy, precision, recall, and f-score are used to evaluate the sparse representation performance on various kinds of existing methods.