It is difficult to use conventional nondestructive testing methods to detect damage, such as loosening of bolted connections, in a space frame structure due to the complexity of the structure and the nature of the possible damage. A vibration-based method that uses changes in the natural frequencies of a structure to detect the locations and extent of damage in it has the advantage of being able to detect various types of damage in the structure, including loosening of bolted connections. Since the vibration-based method is model-based, applying it to a space frame structure with L-shaped beams and bolted joints will face challenges ranging from the development of an accurate dynamic model of the structure to that of a robust damage detection algorithm for a severely underdetermined, nonlinear least-square problem under the effects of relatively large modeling error and measurement noise. With the development of modeling techniques for fillets in thin-walled beams (He and Zhu, 2009, “Modeling of Fillets in Thin-Walled Beams Using Shell/Plate and Beam Finite Elements,” ASME J. Vib. Acoust., 131 (5), p. 051002) and bolted joints (He and Zhu, 2011, “Finite Element Modeling of Structures With L-shaped Beams and Bolted Joints,” ASME J. Vib. Acoust., 131(1), p. 011010) by the authors, accurate physics-based models of space frame structures can be developed with a reasonable model size. A new damage detection algorithm that uses a trust-region search strategy combined with a logistic function transformation is developed to improve the robustness of the vibration-based damage detection method. The new algorithm can ensure global convergence of the iterations and minimize the effects of modeling error and measurement noise. The damage detection method developed is experimentally validated on an aluminum three-bay space frame structure with L-shaped beams and bolted joints. Three types of introduced damage, including joint damage, member damage, and boundary damage, were successfully detected. In the numerical simulation where there are no modeling error and measurement noise, the almost exact locations and extent of damage can be detected.