Optimizing Small-Scale Language Models for Robotic Task Planning: Insights into Prompt Design

Large-scale language models (LLMs) have demonstrated significant potential in robotic task planning, however, their resource-intensive nature poses challenges for on-device deployment in resource-constrained environments, such as autonomous robots. In this paper, we explore the feasibility of utilizing Small-Scale Language Models(SSLMs) for robotic task planning, focusing on the vital roles of prompt design. Through extensive experimentation, we discovered that prompt strategies effective in LLMs are often unsuitable for smaller models. To address this, we developed a set of methods tailored for SSLMs, which include simplifying prompts, decomposing complex tasks, and managing contextual information effectively with reasonable explanations. Our results show that with these prompt design optimizations, SSLMs can perform competitively, providing efficient and reliable task-planning capabilities while operating within the limited computational resources available for robot task planning. This work provides valuable insights and practical guidance for deploying SSLMs in real-world robotic applications.