Physically Consistent Identification of Inertial Parameters for Robotic Arms Using Gradient-based Optimization

Abstract

This thesis presents a physically consistent inertial parameter identification method for robotic arms. The approach introduces an alternative parameterization that guarantees positive masses and density-realizable inertia matrices by construction, instead of enforcing physical consistency through constraints. A reformulation of the inverse dynamics model is used to derive an analytical gradient with respect to the reparameterized inertial parameters, enabling gradient-based optimization. The Franka Research 3 robot is used as the evaluation platform in both simulation and real-robot experiments. In both cases, the experiments are repeated over multiple Monte Carlo trials to evaluate how the proposed method performs under randomized conditions. The simulation experiments analyze convergence behavior, robustness, the influence of the training trajectory and torque-noise sensitivity. The real-robot experiments evaluate whether the proposed algorithm can fine-tune CAD inertial parameters and adapt to changes in the robot dynamics caused by an unknown payload. A comparison with a physically consistent identification method is also presented.