Nonlinear Parameter Identification of the Franka Emika PANDA Robot

Abstract

This research addresses the problem of dynamic parameter identification for robot manipulators. As the complex manipulation of tasks increases, traditional control methods become insufficient, necessitating accurate model-based control. Previous studies have explored various parameter identification methods for robot manipulators. Still, the impact of different friction models on dynamic parameter identification, particularly for the Franka Emika PANDA robot, is yet to be comprehensively investigated. The linear and nonlinear parameter identification methods and nonlinear friction models (Lugre, Stribeck, and Sigmoidal) were studied on the PANDA robot. A linear inverse dynamic model was developed using the Newton-Euler method. Identification and validation trajectories were designed. Nonlinear constraints and bounds were applied for easier convergence, ensuring a positive definite inertia tensor in the center of mass (COM) frame. Using constrained least squares with Lasso regularisation, model-estimated torques comparison to measurements was done to estimate the physical parameters, considering the PANDA robot’s hand as a payload. We obtained the PANDA robot’s physical parameters, including the hand’s mass, without a friction model where discrepancies in joint torques were observed. Incorporating friction models reduced these discrepancies, validated by another trajectory comparison between the measured and the model’s predicted joint torques. Stribeck friction model exhibited the best performance in estimating parameters from its Root Mean Square Error (RMSE) and Mean Absolute Error (MAE), indicating its effectiveness in capturing nonlinearities in the PANDA robot.