Visual Language Models Train Robots to Read Human Emotions

IEEE Spectrum · 4 months ago · 15

Researchers at Monash University trained a vision language model based on Gemini 2.5 to help robots recognize human emotions from facial expressions and contextual cues during collaborative tasks. The VLM achieved an emotion recognition score of 0.86 compared to 0.77 for conventional facial analysis systems, and 31 of 40 human participants preferred emotionally adaptive robot apologies over scripted ones. However, the study found that robots' emotional capabilities matter far less to humans than actual task competence, as participants' trust decreased regardless of how well the robot apologized after failing its physical responsibilities.