Common Sense Plan Verification with Large Language Models

Аннотация

Verifying the plan before executing it on the robot is an important task because a wrong plan can lead to fatal consequences and damage. In this paper, we propose an approach for verifying a plan before executing it on a robot, which combines the use of a formal language and the use of LLM. We first translate the initial instruction, which may not contain all the necessary actions of the plan, into an LTL formula. And then, based on this formula, we complement the plan with a query to the LLM about the possibility of transitioning from one state to another is carried out. We demonstrate the proposed approach on two datasets, a specialized one dedicated to the task of cooking and a general one dedicated to household chores.