The fusion of large language models and social robotics is one of the field’s most significant developments of the last two years. Robots that once depended on structured dialogues can now converse more naturally and respond to situations not anticipated in their design.

What LLMs really improve

The clearest improvement is in open-ended dialogue: robots equipped with LLMs can sustain longer conversations, answer questions outside their expected scope and adapt tone and vocabulary to the person they are talking to. The second gain is in planning complex tasks.

What LLMs do not solve

LLMs do not guarantee physical safety, reliable perception, real emotional understanding, robust manipulation or privacy. A robot that “converses well” still needs safe robotic controllers, reliable sensors and emergency-stop mechanisms.

The architecture that is emerging

The more responsible design is layered: a conversational model for language; a multimodal perceptual model for vision and audio; a planning model for goals; safe robotic controllers for physical action; and an explicit governance layer for ethics and consent.