This thesis describes the design and implementation of an EEG-based motor imagery system for robot control using a 3D printed headset and electrodes. The primary aim was to create a more comfortable and user-friendly EEG headset that, in combination with a deep learning model, can reliably measure EEG signal and classify motor imagery for controlling a robot arm.

Research in this area involves 3D scanning and printing a human head to design and fabricate a custom EEG headset with integrated detachable electrodes. Different electrode materials and coatings were evaluated to determine which ones were most suitable for EEG signal measurement compared with commercial electrodes. Furthermore, machine learning models for binary EEG signal classification using CNNs and transfer learning were developed.

The trained model with best accuracy was then integrated with ROS MoveIt package for controlling a robot arm using user's motor imagery EEG signals. Results showed that the developed EEG headset and electrodes provided reliable, accurate EEG signal measurements for robot control. CNN models achieved high classification accuracy of 93% on public dataset, but poor generalization on personal dataset. Transfer learning provided similar accuracy in comparison with the models trained on public dataset while significantly improved the performance of the model on personal dataset.

In overall, the best CNN model achieved average accuracy rate of 62.5% when testing was made with EEG data obtained in different environments. By connecting the machine learning model to ROS MoveIt package, specific predefined movements can be executed based on user's motor imagery EEG signals. Overall, this research presents a promising path towards creating more comfortable and effective EEG-based robot control systems.