Satellite-Assisted 6G Wide-Area Edge Intelligence: Dynamics-Aware Task Offloading and Resource Allocation for Remote IoT Services

Background Introduction

With the advent of the 6G mobile communication network, the traditional Internet of Things (IoT) architecture is gradually transforming into the new paradigm of the intelligent Internet of Everything (IoE), which integrates global connectivity and extensive artificial intelligence (AI) capabilities. However, terrestrial networks are limited in coverage, especially in complex terrains and remote areas where full coverage cannot be achieved. The rapid development of Low Earth Orbit (LEO) satellite technology brings new hope to address this issue. With the support of Non-Terrestrial Networks (NTN), LEO satellites are able to provide seamless connectivity, high-capacity communication, and efficient computing services to global users, meeting the needs of applications such as remote environmental monitoring and intelligent agriculture.

However, wide-area IoT applications pose challenges for handling computation-intensive tasks, especially when real-time responses are required. Traditional methods involve offloading tasks to ground data centers via the satellite network; however, the transmission latency between satellites and data centers significantly increases and cannot meet the real-time requirements of urgent tasks. The introduction of Multi-Access Edge Computing (MEC) offers the possibility of equipping LEO satellites with edge computing servers, thereby shifting computational capabilities to the edge to reduce latency through on-orbit computing. Nonetheless, the highly dynamic spatio-temporal characteristics of the LEO satellite network, such as high mobility of satellite nodes and intermittent communication links, pose significant challenges to task offloading and resource allocation. Efficient optimization of these issues in LEO-enabled IoT networks thus becomes a pressing challenge in academic research.

Against this backdrop, researchers Di Zhao, Rui Ding, and Bin Song conducted this study and published it in Science China Information Sciences. By introducing spatio-temporal attention mechanisms and combining Deep Reinforcement Learning (DRL), the research team proposed the Spatio-Temporal Attention-Based Proximal Policy Optimization (STA-PPO) algorithm to achieve efficient decision-making in highly dynamic environments for task offloading and resource allocation.

---

Study Provenance

The paper, titled Satellite-assisted 6G wide-area edge intelligence: Dynamics-aware task offloading and resource allocation for remote IoT services, was authored by Di Zhao (Xidian University), Rui Ding (China Satellite Network Group Co., Ltd.), and Bin Song (Xidian University). Published in February 2025 in Science China Information Sciences, Volume 68 Issue 2, it was made available online in early January 2025.

The research was supported by China’s National Key Research and Development Program and the National Natural Science Foundation and focuses on cutting-edge studies of IoT and LEO satellite networks in China.

---

Research Methodology and Workflow

System Architecture and Model Design

The researchers proposed a dynamic LEO satellite-supported remote IoT network that integrates a collaborative “cloud-edge-device” architecture comprising four layers:

1. Device Layer: Composed of IoT devices such as sensors, controllers, and actuators, responsible for executing computation-intensive tasks like real-time monitoring or data analytics.
2. Edge Layer: Comprising LEO satellites equipped with flexible payloads and edge computing servers to provide computing capabilities and network access.
3. Cloud Layer: Connecting to the core network via ground stations to support large-scale remote data processing.
4. Service Layer: Supporting multi-type applications, such as environmental monitoring and intelligent agriculture.

The researchers formulated task offloading and resource allocation as a Mixed Integer Non-Linear Programming (MINLP) problem, transformed into a sequential decision-making problem using reinforcement learning, and further defined it as a Markov Decision Process (MDP).

Workflow Details

1. Spatio-Temporal Dynamics Modeling
   The study modeled time dynamics of task arrivals (e.g., Poisson-distributed characteristics of arrival rates) and spatial dynamics of the LEO network (e.g., satellite positions and coverage changes).

2. Problem Decomposition and Model Transformation
   Delays were modeled based on constraints, including:

   • Coverage delays
     
   • Communication delays and link transmission rates
     
   • Computation delays and processing capability distribution for tasks.
     

To address problem complexity, the joint optimization problem was decomposed into sequential decision-making problems in time slots, solved using a DRL-based framework.

1. STA-PPO Algorithm Design

   • Temporal Attention Mechanism (TAM): Analyzes temporal dependencies in task arrivals to dynamically optimize task offloading strategies.
     
   • Spatial Attention Mechanism (SAM): Captures topological relationships in LEO satellites to evaluate state values based on spatial dynamics.
     
   • Multi-Actor-Critic Architecture: Designed for coverage, task offloading, data transmission, and computation, with embedded attention mechanisms.

2. Simulation Experiments
   Simulations were conducted using Starlink satellites with 1,584 satellites uniformly distributed across 72 orbital planes, observing system efficiency from task initiation to completion. Key parameters (e.g., on-board CPU capabilities, IoT device task sizes and frequencies) were precisely designed and analyzed.

---

Experimental Results and Findings

1. System Delay Performance
   The STA-PPO algorithm achieved the lowest average system delay, reducing delay by approximately 66% compared to random baseline algorithms and by over 23% compared to standard PPO algorithms, showcasing its adaptability to dynamic task fluctuations.

2. Optimization of Task Offloading Ratios
   The optimal algorithm reduced the proportion of tasks processed on local devices while significantly increasing the proportion of tasks offloaded to the edge and cloud, successfully alleviating computational pressure on devices.

3. Link Throughput and Network Throughput Improvements
   The proposed algorithm maximized bandwidth resource utilization, significantly improving uplink and downlink throughput. It performed exceptionally well under high task arrival rates (e.g., when λ=7 or 8), with network throughput improved by up to 19%.

4. Robustness in Dynamic Scenarios
   Simulations under various satellite coverage scenarios demonstrated that the STA-PPO algorithm maintained excellent performance even amid large-scale dynamic changes (e.g., frequent satellite topology shifts or task traffic surges). It met the advanced real-time and efficiency requirements of 6G networks.

---

Research Significance and Highlights

1. Novel Algorithm Design
   The STA-PPO algorithm innovatively combines spatio-temporal attention mechanisms with reinforcement learning, providing efficient task offloading and resource allocation strategies for highly dynamic LEO networks.

2. Low-Latency Advantage
   The study successfully addressed the challenge of real-time computation in high-load IoT tasks, significantly enhancing system efficiency.

3. Large-Scale Application Potential
   The findings support the evolution of base station architecture from terrestrial to “satellite-enabled collaboration,” offering robust theoretical support for data processing and network deployment in fields like agricultural monitoring or disaster response in remote environments.

By considering the dynamic characteristics of LEO satellite networks, this study advances wide-area edge intelligence in 6G and offers efficient technical solutions for remote IoT services. Future research may explore integration with UAV trajectory optimization, multi-carrier network collaboration, and other multi-dimensional technologies for more comprehensive space-based edge computing methods.