VEDA: A Multimodal Dual-Attention Framework for Contact-Aware Temporal Anomaly Detection in Industrial Humanoid Manipulation

Industrial humanoid robots are increasingly expected to execute high-precision manipulation tasks, such as assembly, welding, and material handling, under dynamic and contact-rich working conditions. In such scenarios, operational safety and task reliability depend not only on accurate motion execution but also on timely detection of abnormal interaction patterns between the robot and its environment. This paper proposes VEDA (Visuo-Embodied Dual-Attention), a contact-aware multimodal framework for temporal anomaly detection that explicitly couples contact-guided spatial emphasis with stage-sensitive temporal modeling in industrial humanoid robots. VEDA integrates visual, force-aware, and motion-based embodied perception with a spatial–temporal dual-attention modeling mechanism. The spatial attention module emphasizes contact-relevant regions, while the temporal attention module captures critical operation stages and abnormal transitions over time. The core technical contribution lies in a dualattention design in which force- and motion-related embodied cues guide where the model attends in space and when it attends in time during manipulation monitoring. By jointly modeling multimodal sensory observations and temporal dependencies, VEDA improves anomaly recognition, anomaly localization, and response timeliness during complex manipulation processes. Extensive experiments on benchmark manipulation datasets show that VEDA consistently outperforms representative baseline methods in anomaly detection accuracy, temporal precision, and robustness, demonstrating the effectiveness of task-oriented multimodal attention modeling for reliable industrial humanoid operation.