Efficient Emotion-Aware Iconic Gesture Prediction for Robot Co-Speech

In three linesA lightweight transformer predicts robot co-speech iconic gestures from text and emotion alone, without audio at inference time. The model outperforms GPT-4o on semantic gesture placement classification and intensity regression on the BEAT2 dataset, while remaining computationally compact for real-time embodied agent deployment.

Summary generated by Claude — human-verified