Abstract: The application of collaborative robotic arms is steadily increasing across both service and industrial domains. From the perspective of human-robot safety, rapid and accurate estimation of external torques has become a critical prerequisite for effective collision detection and recognition. However, in real-world deployment and operation, deviations between actual and nominal parameters of the robotic arm’s dynamic model lead to significant discrepancies between conventional external torque estimations and the actual values. To address this issue of model-plant mismatch, a sliding mode momentum observer that exhibits high robustness, reduced dependence on noise distribution assumptions, and low sensitivity to model accuracy was proposed. By systematically comparing the advantages and limitations of existing observers, a quadratic sliding-mode momentum observer is proposed, leveraging the quadratic term to balance rapid responsiveness with smooth filtering. This approach enables reliable collision detection and external torque estimation even in the presence of modeling inaccuracies in system dynamics and noise characteristics. The effectiveness and feasibility of the proposed observer are validated through both simulation studies and physical robotic experiments.