Industrial Robot Trajectory Tracking Using Multi-Layer Neural Networks Trained by Iterative Learning Control

Neural-Learning Trajectory Tracking Control of Flexible-Joint Robot Manipulators with Unknown Dynamics

“…3) Iterative Learning Control: We can implement ILC with the trained RNN that emulates the forward dynamics of the manipulator to search for the optimal input for a given desired trajectory. We use the multiple-input multiple-output (MIMO) gradient-based ILC algorithm as developed in [22] based on the work in [26]. The key idea of the algorithm is that, we iteratively update the input u k (k represents the number of iterations) until convergence with the gradient of the tracking error with respect to the current input u k as below:…”

“…For each new desired trajectory, ILC must be re-implemented to search for the optimal input. In our work [22], we trained multi-layer feedforward NNs offline to find a good estimate of the inverse dynamics of an industrial ABB robot for trajectory tracking control, by implementing ILC for a large number of desired trajectories to collect training data in a high-fidelity simulator. We also applied transfer learning to transfer the learned representations in simulator to the real-world by fine-tuning the NNs with real-world data.…”

“…1) Data Collection: Instead of collecting training data through ILC as we did in [22], we can use the same collected training data above but use the output trajectory q as input and the desired trajectory q d as output to the BRNN, inspired by [19]. The idea behind this is that, if q d and q are corresponding input and output trajectories for a dynamical system G, then with input of q, the inverse dynamical system would produce output of q d .…”

“…Also, the second approach works slightly better than the first method. One possible reason is the initial transient process at the beginning when using ILC [22].…”