Final Project for the course in Robotics 2 A.Y. 2019/2020.
In order to make a robot system execute tasks as regulation and trajectory tracking it needs definitely a feedback control which effectively correct the errors of the feed-forward commands. Standard control methods consist in the inversion of the dynamics model in order to calculate the necessary torque that minimizes the error between the robot reference configuration and the actual one. A way more precise, energy-efficient and suitable for complaint robots control method is the so-called Computed Torque which corresponds to use the dynamic model over the system considering the joint weakly coupled. However, this method requires to exactly know the dynamic robot model and therefore it is not an optimal neither a robust control method. Therefore almost always we have some mismatch between the dynamic model (the nominal one) and the actual behavior of the robot, also after identification procedures. In such cases it is convenient to implement robot learning algorithms in order to better approximate the dynamic model and being able apply a correction to the Feedback Linearization (FL) and so correctly linearize the system. In particular, in this project, we are going to implement the Gaussian Process Regression over the corresponding torque errors between the nominal dynamic models and the ”real one”.
A detailed report of the project
The presentation we gave
MATLAB Robotics System Toolbox
A. Mauro
G. Fioretti
E. Nicotra