On the Role of ANN on the Estimation of Payload Inertia in Parallel Kinematic Machine

In the robotic field, accurately estimating the inertial parameters of a robot payload is essential for improving path-following precision in model-based robot control. Nowadays, there are two main approaches to address this problem. One approach involves using direct measurement of the interaction between the robot End-Effector and the payload in terms of motions and exchanged forces-torques. The other approach involves a standard dynamic robot identification problem, considering the payload as an extension of the last robot link. Several of these techniques have been tested and consolidated for serial robots. However, results obtained for the estimation using Parallel Kinematic Machines have been less accurate than those using serial robots, especially when no direct measurement is considered, due to the complexity of their kinematic chain and the greater impact of uncertainties in their geometrical and dynamic parameters. The purpose of this work is to evaluate the performance of an approach for estimating the inertial properties of the payload using a trained data-driven model, namely an Artificial Neural Network (ANN). This model takes as input the same information available to the machine when direct measurements are made (i.e., positions and forces exchanged between the End-Effector and the payload) and returns as output the inertia tensor and center of mass of the payload. The developed method is validated and compared with standard model-based approaches (e.g., Least Square Method based on the Newton-Euler equations). This work is the first step towards developing a novel approach for estimating the inertial properties of the robot payload using data-driven models with the inputs of the indirect method (i.e., actuator's motion and torque), in order to use as few sensors as possible.