A reasoning-LLM-based high-level planning framework for robotic shelf-stocking and disposal tasks

Labor shortages in the retail industry, particularly in convenience stores, pose serious challenges for shelf-related tasks such as stocking and disposal. This paper formulates shelf-stocking and disposal as a partially observed sequential decision-making problem in which a robot must infer a structured shelf state and generate feasible high-level actions under task constraints. We propose an automated planning framework that combines ArUco-marker-based shelf-state recognition, homography-based discretization, prompt generation using static task rules and dynamic shelf states, large-foundation-model-based action planning, plan validation, and reobservation-based feedback. The generated actions are not executed directly; instead, they are parsed and checked against task constraints before being sent to robot execution primitives. Simulator experiments evaluate plan efficiency and adaptability to complex initial states, while real-robot experiments evaluate task completion and failure modes under physical uncertainty. The results indicate that the proposed framework can generate executable high-level plans and recover from some execution failures through reobservation, while perception errors and model-output uncertainty remain important limitations for practical deployment.