Hof, Lars (2026) Hamiltonian Neural Networks for System Identification and Control of Robotic Systems. Research Project, System and Control.
|
Text
MasterResearchProject.pdf Download (11MB) | Preview |
Abstract
Accurate dynamic models are essential for robotic control but hard to obtain due to nonlinearities, friction, and sensor limitations. This thesis compares classical and learning-based identification methods for a 2-DOF manipulator and evaluates model-based controllers experimentally. Three methods are tested on a Quanser 2-DOF platform: IDIM-OLS exploiting linearity in inertial parameters, Nonlinear Least Squares (NLS) adding Coulomb friction, and a Hamiltonian Neural Network (HNN) parameterizing dynamics within a port-Hamiltonian framework, guaranteeing passivity by construction. HNN sub-networks are warm-started from least-squares estimates to avoid meaningless local minima. NLS achieves the best torque reconstruction; the HNN yields the lowest simulation error with more balanced cross-joint predictions. Warm-starting matches IDIM-OLS accuracy before end-to-end training. Three velocity-free controllers are compared: port-Hamiltonian with momentum observer, discontinuous output-feedback, and finite-time output-feedback. The first two achieve the best tracking. Under disturbances, all show linear error growth. The port-Hamiltonian and finite-time controllers produce smoother torques; the discontinuous controller saturates actuators. A custom calibration station identifies DC motor-gearbox parameters. All experiments run on real hardware at 1 kHz.
| Item Type: | Thesis (Research Project) |
|---|---|
| Supervisor name: | Ahmed, S. and Wilhelm, E. |
| Degree programme: | System and Control |
| Thesis type: | Research Project |
| Language: | English |
| Date Deposited: | 27 Mar 2026 10:20 |
| Last Modified: | 27 Mar 2026 10:20 |
| URI: | https://fse.studenttheses.ub.rug.nl/id/eprint/37267 |
Actions (login required)
 |
View Item |