Joint torques estimation of a robotic arm using neural networks

The identification of accurate dynamic models is still a challenging task and its knowledge is required in a large variety of robotics applications, for example in tasks requiring the detection and estimation of external forces, in high-precision control of movements and trajectories or in tasks involving the interaction with the external environment or with human operators. In order to derive these models, several techniques have been explored in literature. Traditional model-based approaches derive them directly from the laws of Lagrangian mechanics. These strategies, however, require a precise knowledge of the robot parameters, as well as the ability to analytically describe all the dynamical behaviors involved. The presence of parameter uncertainties or unconsidered dynamics, such as complex motor frictions or joint flexibilities, strongly limits their performance. To overcome these issues, purely black-box techniques have recently attracted researchers’ attention. These methods learn the models directly from experimental data, without needing any prior knowledge about either the system parameters or the involved dynamics. In particular, in this thesis project the joint torques of an industrial robotic arm with 6 Degrees of Freedom (DoF) have been estimated by using Neural Networks (NNs). The latter have been arranged in three different architectures: single NN, multiple NNs and cascade NN. The implemented models have been tested on the acquired datasets and their performance has been evaluated in terms of test loss according to metrics such as the Mean Square Error (MSE). The experiments have highlighted the effectiveness of a pre-processing phase of the input dataset and of exploiting the dependencies between joint torques. Finally, the work carried out has shown that Machine Learning (ML) methods such as NNs allow to avoid the complex computations of traditional analytical approaches, by exploiting ML and obtaining good estimates even on complex non-linear models, such as 6 DoF manipulators.