<![CDATA[For millions of deaf-mute individuals, sign language is the only means of communication; this creates barriers in daily interactions with non-signers, leading to the exclusion of these individuals in many areas of daily life. To address this, we propose a real-time sign language translation system using a Transformer model enhanced with a knowledge graph, designed for Human-Robot Interaction (HRI) with NAO robots. Our system bridges the communication gap by translating gestures into natural language (text). We used the RWTH-PHOENIX-Weather 2014T dataset for initial training, achieving a BLEU score of 29.1 and a Word Error Rate (WER) of 18.2% surpassing the baseline model. Due to the domain shift between human gestures and NAO robot gestures, we created a NAO-specific dataset and fine-tuned the model using transfer learning to accommodate an adapted environment and kinematic constraints that do not match the environment in which the robot was deployed. This reduced the WER to 17.6% and increased the BLEU score to 29.9. We tested our model’s capability with dynamic and practical HRI scenarios through comparative experiments in Webots. Integrating a knowledge graph into our model improved contextual disambiguation, significantly enhancing translation accuracy for gestures that weren't clear. Through effectively translating gestures into natural language, our system demonstrates strong potential for practical robotic applications that promote social accessibility.]]>
