On-line expectation-based novelty detection for mobile robots

摘要

This paper presents a recurrent neural network based novelty filter where a Scitos G5 mobile robot explored the environment and built dynamic models of observed sensory-motor values, then the acquired models of normality are used to predict the expected future values of sensory-motor inputs during patrol. Novelties could be detected whenever the prediction error between models-predicted values and actual observed values exceeded a local novelty threshold. The network is trained on-line; it grows by inserting new nodes when abnormal observation is perceived from the environment; and also shrinks when the learned information is not necessary anymore. In addition, the network is also capable of learning region-specific novelty thresholds on-line continuously.