CONTEXT-AWARE DEEP LEARNING FRAMEWORK FOR EMOTION RECOG-NITION AND INTENT PREDICTION IN HEALTHCARE ROBOTS

Abstract

The implementation of intelligent robotic technologies in assistive devices and elderly-care environments creates an opportunity to develop emotion-aware robots that provide timely, personalized support. However, affective relational interaction in assistive robotics remains underdeveloped, particularly where facial, vocal, linguistic, and conversational evidence must be interpreted together. To address this concern, this study develops a multimodal deep-learning framework for emotion recognition and emotion-conditioned intent prediction. Synchronized audio, video, and text from IEMOCAP are processed by wav2vec 2.0-BiLSTM-attention, ResNet-50-BiLSTM, and BERT-BiLSTM-attention branches, respectively. Their representations are concatenated and transformed by dense layers with dropout and Softmax classification. The implemented experiment uses four emotions (angry, happy, neutral, and sad), an 80:20 train-test split, Adam optimization at a learning rate of 0.001, batch size 16, 200 epochs, dropout 0.5, and categorical cross-entropy. Performance is recomputed directly from the reported confusion matrix: 703 of 898 samples are correct, giving 78.29% accuracy, macro-precision 77.02%, macro-recall 79.09%, and macro-F1 77.68%. Sadness obtains the highest recall (89.86%), whereas happiness has the lowest F1 (69.53%). The framework also conditions the proposed intent representation on recognized emotion; however, the available experiment does not provide an intent taxonomy or intent-specific ground truth and metrics. The evidence therefore validates the multimodal emotion component and establishes an architecture for future joint emotion-intent evaluation, while healthcare deployment, response latency, and clinical effectiveness remain outside the demonstrated scope.