This study evaluated whether machine learning models trained to identify arm movements in controlled laboratory settings could accurately recognize those same movements when people performed natural activities in a kitchen environment. The researchers had twelve healthy participants perform three types of arm movements (reaching, lifting, and wrist rotation) using wearable motion sensors. First in structured, controlled trials, and then in semi-structured kitchen tasks like preparing a meal. Both the Random Forest classifier and the hybrid deep learning model maintained strong performance when generalizing from the structured to the semi-structured environment, achieving 86.54% and 87.96% balanced accuracy respectively for subject-specific models, and 77.37% and 82.96% for group models that could work with new individuals' data. The models could even identify movements pretty well using just a single wrist sensor rather than four sensors along the arm, though with reduced accuracy, the Random Forest model maintained approximately 76% accuracy with one sensor in the subject-specific approach. The discussion emphasizes that these findings demonstrate practical feasibility for home-based stroke rehabilitation monitoring, as the models can generalize across different settings without requiring extensive patient-specific training data or burdensome sensor configurations.