Robot learning techniques for set point regulation of robots with nonlinear flexibility on the joints

This is our Final Project for the course in Underactuated Robotics A.Y. 2020/2021.

The task has been to combine learning techniques with Nonlinear MPC in order to perform set-point regulation of a manipulator with elastic joints. In particular, the elasticity term is assumed to be nonlinear and unknown (hence the need of the learning techniques).

Project structure

Here we summarize the project structure for ease of use.

• simulation.m: main file of the project
• simulationOnline.m: main file of the project used to run the simulations with online retraining
• parameters.m: collects all the parameters across the project
• robotModel.m: derives symbolically (and saves numerically as MATLAB functions) the needed terms of the robot dynamical model
• mpcSetup.m: sets up the MPC object
• plotResults.m: displays the results of the simulations

Some utility functions are used.

• dataGeneration: collects all the utilities for generating offline training data
• modelFunctions: collects the functions for the terms of the robot dynamical model
• modelsTraining: collects all the utilities for training the learning models
• mpcFunctions: collects all the functions for MPC (prediction model, custom cost function, custom constraints)
• savedData: collects generated datasets and trained models
• utils: collects various utility functions used across the project

Documentation

Check the report of the project here and the presentation we gave of our project.

Resources used

• MATLAB Symbolic Math Toolbox
• MATLAB Statistics and Machine Learning Toolbox
• MATLAB Model Predictive Control Toolbox
• CoppeliaSim

Authors

• Andrea Caciolai
• Emanuele Nicotra
• Matilde Rabbiolo