Robots today still struggle with adaptation and generalization to changes in the task, a major barrier to deploying robots in semi- and unstructured tasks. If robots can detect novel situations that are out-of-distribution, defined as scenarios not represented in the training data, they can initiate a fail-safe or fallback action to recover or at least avoid damage. Anomaly detection identifies data patterns that deviate from expected behavior. We apply a variational autoencoder (VAE) approach to time series in robotics for an industrial cabling task. Inputs are force measurements and the robot’s end-effector positions, from both nominal processes and various failure scenarios. In validation, the VAE model achieved an AUROC of 0.82 in detecting process-related failure. In the overall evaluation, two of three types of failures were reliably detected, while the third, which had smaller magnitude deviations in the force profile, proved challenging to identify robustly.