Since Hans Berger's 1920s introduction of EEG, its signals have been vital for diagnosing neurological issues and other applications. Brain-computer interfaces (BCIs) have transformed mind-controlled robotics, with Deep Learning enhancing EEG signal decoding. However, factors like substance abuse and Parkinson's disease pose classification challenges. To address these complexities, our research designed and validated a dual-model architecture. The aim was to generalize motor imagery classification for mind-controlled robots, encompassing both healthy individuals and those affected by substance addiction. This extended to assessing the real-world applications of these models in piloting advanced robotic platforms, including quad-rotors and wheeled mobile robots. Our empirical evaluation highlighted clear distinctions between the SVM and attention-based Bi-LSTM models; With the SVM, precision values fluctuated within a range of 0.702 to 0.782, while recall metrics varied between 0.705 and 0.790. On the other hand, the attention-based Bi-LSTM model demonstrated a broader precision range of 0.744 to 0.864 and recall values spanning from 0.752 to 0.856. Impressively, the average precision and recall for the Bi-LSTM model stood at 0.807 and 0.809 respectively, indicating its robust and consistent performance. Additionally, the presented confusion matrix further substantiates the heightened efficiency of the attention-based Bi-LSTM model in EEG signal classification. The evident proficiency of this model in classifying EEG signals ushers in an innovative and inclusive approach to mind-controlled tasks. This not only advances the BCI field but also signals promising avenues for therapeutic and rehabilitative measures tailored for those with neurological discrepancies.