A Visual Dataset for Anomaly Detection in Self-Driving Laboratories

Self-driving laboratories accelerate the application of the scientific method and the discovery process through high-throughput experimentation, intelligent perception and planning, and effective human-robot collaboration. However, detecting anomalies in events, object states, and environmental conditions remains challenging due to process uncertainty and environmental complexity. To support research in this area, we construct a dataset for process anomaly detection in scientific experiments. The dataset is collected from a fully automated Polydimethylsiloxane synthesis workflow involving collaborative robots. Images were captured from first-person views using end-effector cameras mounted on mobile and fixed robotic arms, covering 11 checkpoints and 14 distinct viewpoints. In total, the dataset includes 1,671 images and 2,788 image-text pairs. Each sample contains step-specific descriptions, anomaly labels, and region-level annotations. This dataset intended to support a range of visual anomaly detection tasks, including image-text classification, anomaly type recognition, anomaly localization, grounded captioning, and anomaly attribution analysis. It offers a practical resource for developing intelligent anomaly monitoring and automated decision-making systems in self-driving laboratory environments driven by visual perception.