This repo contains algorithms for Model Predictive Control, ensuring safety constraints and recursive feasibility. It is a general framework that can work with any manipulator. However for each system and environment (customizable in config.yaml) a network that approximate a control invariant set is needed.
Until now we tested the framework with the Unitree Z1 robot. URDF files and networks for the sets for some robot-environments settings can be downloaded here. Place the folder robots and nn_models in the root folder and follow the instruction in nn_models to use the networks properly.
- Clone the repository
git clone https://github.com/idra-lab/safe-mpc.git - Move on the branch devel
- Install the requirements
pip install -r requirements.txt - Follow the instructions to install CasADi, Acados and Pytorch.
- In root folder, create a folder called robots, and place in it the needed subfolders containing the URDF files of the robots
- Create also a folder named nn_models, in which are located the networks approximating the control invariant sets.
- In config.yaml, set the simulation parameters
- Run guess_acados.py (-- help to get instructions about the line command arguments) to find the initial configuration and the warm start trajectories. In guess_acados set also the type of cost and task defining cost_controller, and choosing one of the classes defined in cost_definition.
- Run mpc.py to execute the RTI MPC control. Also here set the appropriate cost_controller.
- If you want to apply model noise in the simulation, generate the perturbed URDF files running generate_urdf_noise.py and specifying the amounts of noise desired. Then compute the guess as usual with guess_acados.py and run the simulation using mpc.py. To get information regard the use of noise parameters, use the option --help on mpc.py or guess_acados.py.
