Model-based, torque-level control can offer precision and speed advantages over velocity or position-level robot control. However, the dynamic parameters of the robot must be identified accurately. Several steps are involved in dynamic parameter identification, including modeling the system dynamics, joint position/torque data acquisition and filtering, experimental design, dynamic parameters estimation and validation.
The proposed parameter identification process
Our approach employs the inverse dynamic model and least squares (LS) estimation method to estimate inertia parameters of robot arm. We also use a zero-phase low pass filter to process position data and velocities are calculated with a central difference algorithm. Accelerations are calculated with the central difference algorithm and followed by smoothing, which is performed by the Robust LOcal polynomial regrESSion (RLOESS) smoother.
Experiment design usually consists of two steps. The first step is trajectory parameter selection and the second step is parameter optimization. We propose a novel, modified Fourier series, which can generate a persistent excitation trajectory at a greatly reduced level of complexity and computation time necessary for the optimization process.
In this project we developed a novel, simple and intuitive optimal criterion to design the exciting trajectory for the robot to follow. Our approach is able to reduce the number of terms that define the trajectory and employ Hadamardâ€™s inequality to greatly simplify the optimization problem. Indeed, our approach can solve the nonlinear optimization problem in less than one minute with 625 samples, when existing methods need over around 10 minutes. The proposed excitation trajectories also are zero or close to zero at the start and end points.
Experiments with a Staubli TX90 industrial arm show excellent ability to predict acceleration of the joints using the inertial parameters in the dynamic equations. The dynamic parameters of a 6 DOF Staubli TX-90 robot were accurately identified with respect to the proposed excitation trajectory. After discarding the dynamic parameters with a large RSD, we keep 29 essential dynamic parameters to describe the dynamic model of the Staubli TX-90 robot. To validate the reliability of the dynamic parameters, we ran the robot with the rajectories, which were used to estimate the inertial parameters. We also ran the robot with new trajectories, which were not used in the parameter estimation, and compared the measured and the predicted torques. RMS errors were small, with magnitude similar to those reported by other recent researchers.
Jingfu Jin and Nicholas Gans, “Parameter Identification for Industrial Robots With a Fast And Robust Trajectory Design Approach”, Robotics and Computer-Integrated Manufacturing, 31, 2015, pp. 21-29