ST-HMR: A Robust Collision Detection Framework for Collaborative Robots Based on Dual-Domain Interaction and Manifold Regularization

As physical human–robot interaction (pHRI) becomes increasingly prevalent in unstructured industrial environments, ensuring contact safety under complex dynamics remains a critical challenge. External sensors can detect contact effectively, but their cost and integration complexity limit industrial scalability. Sensorless approaches are more economical, yet many data-driven methods ignore frequency-domain mechanical noise, rely on single-scale temporal modeling, and remain vulnerable to unknown anomalies. To address these issues, this article proposes ST-HMR, a sensorless collision-detection framework that relies exclusively on intrinsic proprioceptive signals, including joint angles, joint velocities, and measured joint motor currents. The framework is organized around three interrelated components: ST-GIM derives frequency-aware gates from each proprioceptive window to suppress drivetrain noise, dilated temporal pyramid aggregation aggregates short- and long-range temporal responses for both abrupt impacts and compliant contacts, and hypersphere manifold regularization constructs a compact normal-operation manifold for robust anomaly scoring. Experiments on a UR5e robot show that ST-HMR achieves 94.75% accuracy and an AUC of 0.9814 while maintaining specificity above 96%, providing a practical sensorless solution for industrial robot safety.