Large Language Model (LLM) agents are seen as a promising route to manage long‑term physical tasks without human supervision. Physical tasks demand continuous environment perception, decisive actions, and robustness to changing conditions. Existing approaches either require massive data and retraining or focus mainly on agents operating in virtual worlds.
This paper investigates the feasibility of building a self‑adaptive physical AI agent that can handle long‑horizon tasks in a zero‑shot manner and adapt to environmental shifts without human intervention. We introduce a multi‑agent framework that tightly integrates planning, tool calling, observation, and verification, forming a closed‑loop control pipeline.
Experiments are conducted on agricultural scenarios under varied weather patterns (clear, rainy, stormy) and benchmarked against reinforcement learning (RL) agents. Evaluation metrics include task completion rate, resource efficiency, and adaptability.
Results show that zero‑shot LLM agents achieve comparable management performance to RL agents under identical weather conditions, while exhibiting significantly faster adaptation and higher final performance when the weather shifts, highlighting their superior self‑adaptivity.
These findings demonstrate that LLM agents equipped with planning, tool invocation, and real‑time observation can manage long‑term physical tasks in a zero‑shot setting, offering a promising direction for future self‑adaptive physical AI research. Review