Accurate dynamic models of industrial robot manipulators are essential for achieving high-performance control and for realizing accurate physics simulation. However, improving model fidelity typically requires extensive experimental data and time-consuming identification procedures. This thesis addresses this trade-off by advancing system identification methods aimed at improving the accuracy of physically parameterized robot models. The advancements are focused on resource- efficiency in terms of shortening experiment duration for data collection, decreasing data volume and reducing need for manual effort related to experiment design and tuning of the identification algorithm.

The work first investigates high-fidelity modeling approaches, including multibody formulations with flexible links derived from Finite Element data. These models enable accurate prediction of robot dynamics early in the development process without requiring experimentally identified stiffness parameters. To further enhance model realism, extended joint models incorporating gear, bearing, and link flexibility are developed, allowing separation and identification of transmission- and arm-side dynamics using dual encoder measurements.

The second central contribution of the thesis is the development of improved experiment design strategies for frequency-domain identification. By formulating and solving optimization problems based on information criteria, optimal robot configurations for data collection are identified, significantly reducing experimental effort while maintaining or improving parameter estimation accuracy. These methods are validated both in simulation and on real industrial manipulators, demonstrating substantial reductions in experiment time. The approach is further extended to systems incorporating nonlinear transmission stiffness through the use of statistical linearization.

Third, the thesis introduces advanced estimation techniques for extracting more information from limited data. In particular, local parametric methods for frequency response function estimation are adapted to nonlinear, closed-loop, multi-input multi-output robotic systems, reducing the number of required experiments and improving estimation accuracy compared to classical approaches.

A key contribution of the thesis is the experimental validation of the proposed methods and models, ensuring their practical relevance. Overall, the results demonstrate that substantial improvements in model fidelity can be achieved without increasing experimental resources. By combining physically parameterized models, optimized experiment design, and efficient estimation methods, the proposed framework has significant practical and industrial impact, supporting the development of accurate and computationally efficient dynamic models for industrial robot manipulators.