Modern robotic systems require actuators capable of delivering high precision and reliable performance across diverse tasks. In this context, the BEAR (Belt Elastic Actuator for Robotics), developed by Smile.Tech, is a belt-driven elastic robotic actuator, conceived as an active joint suitable for multiple robotic architectures and applications. Its elastic transmission provides reduced mechanical backlash compared to traditional rigid-gear solutions, ensuring consistent motion. However, the BEAR is currently operated through trial-and-error using position feedback from two encoders—one on the motor and one on the output shaft— whereby performance depends on empirical tuning, reducing reliability in more dynamic operating regimes. This paper presents the identification of the BEAR actuator, producing models that describe its dynamic behaviour. These models make it possible to replace empirical tuning, enabling predictive simulation, supporting advanced control methods, and allowing systematic use of the actuator’s elastic properties. The proposed study relies on dedicated experimental input–output data to characterise the dynamic behaviour of the BEAR actuator under relevant operating conditions. These data are used in grey-box modelling, combining the physical knowledge of the system and data-driven parameter estimation. Prior to application on the BEAR, the approach is validated on progressively complex systems, allowing refinement of the identification procedure and ensuring the reliability of the resulting models. This study aims to develop a consistent model of the BEAR actuator, establishing a quantitative foundation for subsequent analysis and control. By enabling model-based operation and control design, this work seeks to support improved precision, predictability, and reliability in robotic tasks, with potential implications for multifunctional elastic joint applications.