This study aims to optimize real-time detection of upper-limb motor intentions by integrating electroencephalographic (EEG), surface electromyographic (sEMG), and inertial measurement unit (IMU) signals into a multimodal system, with future applications in robot-assisted rehabilitation. Expanding on a previous EEG–sEMG fusion approach, we introduced targeted improvements in feature extraction, classifier optimization, modality integration, and signal weighting. All signals were collected using a custom-designed experimental protocol developed to assess intention decoding under controlled yet realistic conditions. EEG features were extracted using common spatial patterns, while sEMG and IMU features were obtained from time-domain and statistical descriptors. Evaluation employed robust cross-validation methods: Leave-One-Subject-Out and Leave-One-Session-Out. The optimized pipeline markedly enhanced classification performance, achieving accuracies of 98.1% (shoulder flexion-extension), 95.1% (elbow flexion-extension), and 88.5% (shoulder rotation). These improvements were statistically significant (p < 0.05), suggesting they are unlikely due to chance. Moreover, latency was reduced to ensure compatibility with real-time control systems.