Active Anomaly Detection for Autonomous Robots: A Benchmark

Autonomous robots require well-trained Anomaly Detection systems to detect unexpected hazardous events in unknown deployment scenarios. Such systems are difficult to train when data is scarce. During deployment, lots of data are collected but it is not apparent how to efficiently and effectively use that data. We propose to use Active Learning to select samples to improve the detection system performance. We benchmark 8 different query strategies, of which 2 are novel, using normalizing flow over image embeddings. While our results show that our approach has the best performance overall, choosing the right query strategy strongly depends on external factors.