Technical Summary: An Agentic AI Framework with Large Language Models and Chain-of-Thought for UAV-Assisted Logistics Scheduling with Mobile Edge Computing

1. Problem Statement

The paper addresses a hybrid scheduling problem within the context of Cloud Manufacturing (CMfg). The core challenge involves the tight coupling of two distinct operational domains:

1. Physical Logistics: Unmanned Aerial Vehicles (UAVs) must collect finished products from geographically distributed manufacturing stations and return them to a central depot.
2. Computational Offloading: Industrial Sensor Devices (ISDs) at these stations generate computational tasks (e.g., quality inspection, predictive maintenance) that must be processed locally, on the UAVs (Mobile Edge Computing - MEC), or offloaded to the cloud via the UAVs.

Key Challenges Identified:

• Modeling Complexity (Challenge I): The mathematical formulation of this joint problem is significantly more complex than traditional single-domain scheduling. It requires capturing how UAV routing decisions create time-varying "service windows" that dictate when MEC services are available. The model must simultaneously balance conflicting goals: logistics efficiency (product collection) and computational resource allocation under strict energy, capacity, and deadline constraints.
• Solution Complexity (Challenge II): The problem is a large-scale combinatorial optimization task that is NP-hard. Logistics and computation decisions cannot be optimized in isolation. Furthermore, as system conditions and user demands evolve, the optimization landscape changes, requiring a solution method that is both efficient and adaptable.

2. Methodology

The authors propose a two-pronged framework to address these challenges: an Agentic AI Framework for problem formulation and a Hierarchical Deep Reinforcement Learning (DRL) approach for problem solving.

A. Agentic AI Framework for Mathematical Formulation

To address the difficulty of manually modeling complex hybrid problems, the paper introduces an agentic AI system that translates natural language descriptions into interpretable mathematical formulations. This system integrates:

• Large Language Models (LLMs): To process user inputs and generate reasoning.
• Retrieval-Augmented Generation (RAG): To ground the LLM in a curated knowledge base containing domain-specific modeling logic, constraints, and objective patterns from literature and technical documentation. This minimizes hallucinations and ensures domain consistency.
• Chain-of-Thought (CoT) Reasoning: To decompose the modeling task into a structured, step-by-step process (defining objectives, deriving routing constraints, deriving offloading constraints, and summarizing notation).
• Agentic Verification: A multi-agent workflow where a "Responder" generates the formulation and a "Verifier" checks each reasoning step against the retrieved context. If a step fails verification, the system triggers a "rethink" loop before proceeding.

B. Hierarchical DRL for Optimization

To solve the formulated NP-hard problem, the authors design a hierarchical framework based on Proximal Policy Optimization (PPO), decomposing the joint problem into two sequential layers:

• Upper Layer (UAV Routing): Formulated as a centralized Markov Decision Process (MDP). It determines station assignments and visiting sequences for multiple UAVs under constraints of payload, flight distance, and energy. The output of this layer defines the service windows (arrival/departure times) and remaining energy for the lower layer.
• Lower Layer (Task Scheduling & Resource Allocation): Also an MDP, operating on a per-slot basis. Given the service windows from the upper layer, it decides for each task:
  • Execution mode (Local, UAV, or Cloud via UAV).
  • Resource allocation (computing and communication bandwidth).
  • The objective is to maximize task completion within deadlines while minimizing resource occupation and backlog.
• Training Strategy: The layers are trained sequentially. The upper layer is trained first; its best policy is then fixed to generate service window data for training the lower layer. This preserves cross-layer coupling while avoiding the instability of simultaneous optimization.

3. Key Contributions

The paper claims the following contributions:

1. Agentic AI for Hybrid Modeling: It is the first work in smart manufacturing to utilize agentic AI (LLMs + RAG + CoT) to formulate hybrid combinatorial optimization problems. This framework enables the decomposition of complex task requirements into structured, interpretable optimization models that capture intricate dependencies between logistics and computation.
2. Interactive Formulation Process: The proposed process leverages RAG to retrieve domain knowledge and CoT to enhance the interpretability and traceability of the generated models. It allows the AI to reason step-by-step about task couplings and constraints, enabling accurate model construction under cloud-edge-device collaboration.
3. Hierarchical DRL Solution: The paper develops a hierarchical DRL approach that decomposes the coupled routing-MEC decision-making into an upper routing layer and a lower task-scheduling layer. This improves training stability while preserving essential cross-layer coordination through routing-induced service availability.

4. Experimental Results

The framework was evaluated in a simulated cloud manufacturing scenario involving two UAVs, six manufacturing stations, and 1,920 randomly generated tasks.

• Agentic AI Performance: Comparisons with a single-agent baseline (LLM + RAG without CoT/Verification) showed that the proposed framework produced formulations with higher semantic fidelity to the domain knowledge. Specifically, the proposed framework correctly interpreted abstract resource occupation penalties, whereas the baseline incorrectly reformulated them as raw resource usage sums.
• Upper-Layer DRL: The PPO-based routing policy demonstrated robust convergence, achieving a product collection rate approaching 100% in the majority of episodes after initial exploration.
• Lower-Layer DRL: The proposed hierarchical PPO approach was compared against a baseline using Advantage Actor-Critic (A2C) for the lower layer.
  • Stability: PPO exhibited more stable performance in the later stages of training compared to A2C, which suffered from sharp reward drops.
  • Deadline Satisfaction: The PPO-based lower layer maintained a 100% deadline satisfaction rate consistently in the last 500 episodes, whereas the A2C baseline experienced occasional drops below 100%.
  • Collection Success: The system achieved full product collection in 99.6% of the last 500 episodes.

5. Significance and Claims

The paper positions this work as a unified framework for both formulating and solving hybrid logistics-computation scheduling problems in Cloud Manufacturing.

• Significance of Formulation: By automating the translation of natural language requirements into rigorous mathematical models, the framework lowers the barrier for engineers to model complex, tightly coupled cyber-physical systems.
• Significance of Solution: The hierarchical DRL approach demonstrates that decomposing complex joint optimization problems can yield stable, high-performance policies that respect the physical constraints of UAVs and the timing constraints of industrial tasks.
• Modest Scope: The authors note that the current work focuses on homogeneous UAVs and specific manufacturing scenarios. They suggest that future work could extend the framework to larger-scale systems, heterogeneous UAV fleets, and more dynamic manufacturing environments. The paper does not claim to solve all scheduling problems but rather provides a specific methodology for the coupled UAV-MEC scenario described.