Positive emotions significantly impact learning outcomes as they foster problem-solving and decision-making, while negative emotions can hinder information processing. Traditional methods for detecting emotions in immersive virtual reality (VR) applications, such as electroencephalography (EEG) and electrocardiography (ECG), are often invasive, expensive, and susceptible to noise. This paper introduces a new, non-invasive approach for classifying emotional valence in Virtual Immersive Learning Environments (VILE) based on analyzing head and hand movements using machine learning algorithms. A modified Self-Assessment Mannequin collected multimodal behavioral data from 30 students. The results show promising accuracy (71%) in distinguishing between positive and negative emotional states. While this method currently focuses on valence classification rather than specific emotion recognition, it paves the way for developing more accessible and less intrusive solutions for emotional detection in VR learning contexts.