Table of contents

Project Overview

This project focused on developing and testing a thrust control system for a sounding rocket. Two controllers were proposed: one based on Error-based Active Disturbance Rejection Control (EADRC) and another using a neural network model combined with nonlinear model predictive control (NMPC). The work was conducted at PUT Rocketlab as part of my co-authored Master’s Thesis. In this post, I summarize the project’s main goals, challenges, and overall approach.

For readers interested in deeper technical details, the complete thesis is available here.

Technologies used

The project focused on comparing attitude controllers for a sounding rocket, with an emphasis on proposing and evaluating controllers for thrust vector control. Two controllers were compared: the first was a cascade controller consisting of a nonlinear PD controller in the outer loop and an EADRC in the inner loop. The second was a NMPC approach. The NMPC used a hybrid neural model that integrates a known analytical part of the rocket’s dynamics with a neural network. Both controllers employed quaternion representation. Initially, the algorithms were tested in a custom Python simulator modeling a 6 degrees of freedom (6 DoF) rocket. Then, they were evaluated in a custom Python simulator for the Hopper test stand. Finally, the controllers were tested under real conditions on the Hopper test stand.

Test stand overview

The “Hopper” test stand is a VTOL (Vertical Take-off and Landing) platform used to experimentally verify control algorithms designed for sounding rockets with thrust vectored by jet vanes. The setup features motors with propellers generating airflow, which is then deflected by jet vanes to control orientation. The test stand focuses solely on thrust vectoring experiments, not aerodynamic interactions, and simultaneously tests altitude and orientation controllers.

Results

Two test campaigns were conducted. The test scenario focused on maintaining vertical attitude during ascent and after reaching the target altitude. Both controllers were tested with additional disturbances caused by a counterweight used to manage the rope during ascent.

The first campaign aimed to test the cascade controller, with additional evaluation of its robustness to external disturbances. The results met expectations, demonstrating the controller’s ability to maintain or return to vertical attitude despite disturbances such as forces applied by a rope or manually. A video of this test is included below.

The second campaign focused on testing the neural NMPC controller. While this controller also tended to converge to the vertical attitude, its robustness was lower compared to the cascade controller. This difference was partly due to the hybrid model training method and controller tuning.