Effects of Model MisMatch on Controller

May 2023 - Jun 2023

What?

Model-based control method suffer from 1) mismatch between modelled and true dynamics 2) unknown or inaccurate state feedback.
An important question is thus: "How much of an effect does model mismatch have on a controller design?"
The project aims to quantify and understand the effects of this mismatch in scenarios where true system dynamics are unknown or imprecisely modeled, using the cart pole swing-up problem as a case study.

Copy of AA273_Presentation

Developed a cart pole simulation with distinct "true" and "model" dynamics.
Implemented an iLQR controller, paired with various state estimation techniques.
Conducted tests with different levels of model mismatch and delays, using various observer designs.
Employed Kullback–Leibler divergence to quantify model mismatch.
Evaluated controller-observer pair performance based on trajectory cost under different mismatch levels and observer configurations.

result of AA273_Presentation

Model mismatch was found to adversely affect the performance of the controller-state estimation system with a non-linear performance degradation with exponential or step-like trends in some cases.
Findings were concluded in the report shown below.

Check out the report below for more detail!