Enhancing Autonomous Task Execution in Social Robots with Large Language Models

The field of Social Robotics focuses on developing autonomous robots that can interact socially and assist with various tasks. However, the design and execution of such robots are complex they need to understand their environment and accurately interpret human-level instructions, and a common issue is the mismatch between the instructions inputted and the robot's actual performance in specific contexts. This study therefore proposes to enhance a Pepper robot's autonomy and natural interaction by enhanced instructions via Large Language Models (LLMs). Our study involves evaluating commercial and open-source LLMs' performance and comparing different prompting strategies. Both automatic and human evaluations were carried out and showed the potential and limitations of an approach guided by LLMs. Results showed that 55 percent of tasks generated passed the automatic runtime execution assessment, and GPT -4 achieved the highest success rate.