Unsupervised Pose Anomaly Detection for Dynamic Robotic Environments

The integration of intelligent robotic systems in dynamic environments is one important challenge in the industrial automation. These environments require the robots to adjust the calculated poses while they are conducting their tasks in the gripping and the post gripping phase. This work introduces a three-stage approach, called the Reduced Unsupervised Reconstruction Anomaly Detection (RURAD). RURAD detects occurring pose anomalies, which allow the robot to react to dynamic events by adjusting the placing pose. The introduced anomaly detector has been trained based on a systematic training parameter tuning approach, in order to identify the optimal training setup. RURAD achieves a state-of-the-art performance on the introduced pose anomaly detection data set.