Anomaly Detection for Dynamic Human-Robot Assembly

Abstract

Human action recognition is one of the key challenges in human-robot collaboration (HRC), especially when the process has multiple valid ways to assemble a product. To address this problem, we developed an anomaly detection framework for the assembly of complex products. We used an Long-Short-Term-Memory (LSTM)-based autoencoder to detect anomalies in human behavior and post-process the output to categorize it as a green or red anomaly. A green anomaly represents a deviation from the intended order but a valid assembly sequence. A red anomaly represents an invalid sequence. In both cases, the worker is guided to complete the assembly process.