Physics-Informed Neural Networks for Learning and Control
This page describes the implementation of a learning-based nonlinear model predictive control (NMPC) using physics-informed neural networks (PINNs), which is experimentally validated with SPONGE. PINNs are trained in PyTorch, and their hyperparameters are optimized via ASHA. One PINN is integrated into a ROS package (sponge_mpc
) and used as model within NMPC. For this purpose, CasADi is used in a ROS service (C++ node) which communicates with Simulink, enabling its use on the real-time hardware. The test-bench software is based on the originally test-bench software. It is therefore recommended that you familiarize yourself with this.
The code for PINN training, hyperparameter optimization and learning-based NMPC with PINNs can be found in the git repository. Also, 13 open-source real-world datasets of SPONGE with five actuators can be found there.
Additional Requirements
- CasADi installed as a source build with IPOPT solver
- json.hpp copied to include folder
Usage
- Set up the test bench following these instructions. The ROS-interface is used, which is explained here
- Dev-PC: Initialize parameters and open Simulink model via
init.m
- Dev-PC: If necessary, modify Simulink model
- Dev-PC: Compile the model by pressing
Ctrl+b
- Dev-PC: Compile ROS-Workspace and copy to RT-PC via
$ ./build.sh && ./sync.sh
- Connect to RT-PC via SSH and run the following commands on RT-PC:
$ sudo /etc/init.d/ethercat start
(start EtherCAT master) and$ ~/app_interface/ros_install/scripts/autostart.sh && tmux attach-session -t app
(start compiled model) - Dev-PC: Launch ROS service for PINN-based MPC via
roslaunch sponge_mpc sponge_mpc.launch
- Dev-PC: Start external mode in Simulink model via
Connect To Target
to visualize/record data or alter settings (such as starting the MPC experiment) - After the experiment on RT-PC:
Ctrl+c
in tmux window,$ tmux kill-session -t app
and$ sudo /etc/init.d/ethercat stop
to stop the EtherCAT master
Citing
The paper is freely available via arXiv. If you use parts of this project for your research, please cite the following publication:
Generalizable and Fast Surrogates: Model Predictive Control of Articulated Soft Robots using Physics-Informed Neural Networks
T.-L. Habich, A. Mohammad, S. F. G. Ehlers, M. Bensch, T. Seel, and M. Schappler
Currently under review