Can only LLMs do Reasoning?: Potential of Small Language Models in Task Planning

In recent robotics, Large Language Models (LLMs) have been noted as domain-agnostic task planners due to their advanced perception and reasoning capabilities. Especially, the reasoning capability of LLMs based on Chain-of-Thought (CoT) is gaining attention for robotic task planning. However, we claim that LLMs are excessive for robotic task planning for two reasons. First, the language commands given to robots have much lower linguistic complexity than what LLMs can handle. Second, modern robots in practice are primarily tailored to a specific domain or environment (i.e., kitchen). Building on this, we further believe that within a single, specific environment, small language models (LMs) also have the potential to be effective robotic task planners. To explore this, we generate ‘COmmand-STeps dataset’ (COST) via LLMs, which is a dataset designed for environment-specific robotic task planning. This dataset is then used to fine-tune small LMs to plan action steps for robots in response to given human commands. The applicability of COST dataset to real-world robots is verified by showing that the action steps generated by fine-tuned small LMs on COST dataset are practically executable by the robot. Furthermore, through comparison between LLMs and small LMs fine-tuned on COST dataset within specific task environments, it is demonstrated that the fine-tuned GPT-2 medium is comparable to GPT-3.5 in environment-specific task planning.