LSTM-based Inverse Dynamics Learning for Franka Emika Robot

This paper investigates long-short-term-memory-based learning of inverse dynamics for an industrial robot from experimental data. A Franka Emika Robot is used as a case study to estimate the inverse dynamic model. An automated dataset acquisition method is presented. Furthermore, a long-short-term-memory-based neural network was built and optimized with Bayesian hyperparameter search subject to hyperband termination. Both Matlab and Python implementations of the robotics toolbox are used for comparison with conventional baseline methods as well as measured ground truth data. The generated dataset possesses near axis-limit mean centered Gaussian distributions of the command motion parameters (joint position, joint velocity and joint acceleration) and the resulting measured joint torques. Accuracy of the proposed learned inverse dynamics model (mean rmse 0.787 Nm) is comparable with state- of-the-art deep-learning-based approaches and conventional non-deep-learning-based approaches.