6G AI-Native Network with Edge-Device Collaboration

1. Background of 6G Edge-Device Intelligent Collaboration Research

The continuous innovation in mobile communication and Artificial Intelligence (AI) technology is driving 6G towards deep integration with sensing, intelligence, computing, and security. This expands network service boundaries and enhances service efficiency, evolving from 5G's "Internet of Everything" to "Internet of Intelligence." 6G services are increasingly geared towards next-generation intelligent terminals with integrated sensing, intelligent decision-making, and computing capabilities. Terminal forms are diversifying beyond traditional mobile phones to AI phones, smart wearables, intelligent robots, and intelligent connected vehicles, promoting the diversification and ubiquity of 6G intelligent services. To meet future ubiquitous intelligent service demands for low latency, high precision, and efficient data processing, the industry proposes strategies to deploy models at the edge or terminal side. Terminals handle local lightweight intelligent tasks, while the edge focuses on complex computing tasks offloaded by terminals, significantly improving intelligent service metrics through edge-device collaborative computing.

Currently, global standardization organizations are actively exploring key technologies for 6G edge-device intelligent collaboration for intelligent terminals. The US NextG Alliance advocates for considering terminal AI computing capabilities in 6G system design, addressing intelligent task collaboration between terminals and network computing nodes to meet the challenges of diverse data processing and resource constraints in IoT devices. The European 6G-IA association includes intelligent collaboration scenarios like collaborative robots in key 6G application scenarios, emphasizing AI-driven air interface intelligence and edge intelligence. Research includes distributed model training and deployment technologies for edge devices, including terminals. China's IMT-2030 (6G) Promotion Group proposes that 6G mobile communication networks will provide ubiquitous AI-as-a-Service (AIaaS) capabilities. Research is being conducted on key technologies for edge-device intelligent collaboration, including edge intelligence and distributed training/inference. For the wireless air interface, research is focused on data formats and model designs integrating mobile communication and AI to provide model references for industry development. In 2025, the 3rd Generation Partnership Project (3GPP) Release 20 officially launched a 6G network architecture research project, designating network AI framework design as a core research area and exploring efficient, scalable data control frameworks.

Academia is actively researching edge-device intelligent collaboration methods for data, model, and resource optimization. Several studies propose methods for model splitting and resource optimization in edge-device collaboration to reduce system energy consumption, jointly optimize model splitting, data compression, and service access strategies based on terminal mobility to reduce service latency, and optimize bandwidth, edge computing power, and task allocation in cloud-edge-terminal computing networks to enhance overall network utility.

2. 6G Edge-Device Intelligent Service Characteristics and Requirements

Edge-device intelligent services are characterized by strong sensing, strong computing, and strong collaboration. They meet multi-dimensional requirements such as low latency, high precision, and high reliability through data fusion processing, task collaboration, and real-time decision-making. Overall, the development of 6G edge-device intelligent services exhibits three major characteristics:

2.1 AI-Native Services

The deep integration of AI technologies like large models and intelligent agents with terminals will foster new application paradigms for AI-native services in 6G [12]. This trend places higher demands on network service provisioning capabilities, shifting service demands from traditional mobile access to intelligent services that combine "connectivity + AI." Consequently, the network must accurately identify the heterogeneous resource needs (e.g., communication, computing) of terminal AI services, enabling intelligent, on-demand integrated services for smart terminals through a unified decision-making mechanism.

2.2 Integrated Sensing

With the increasing autonomy and intelligence of devices like intelligent robots and intelligent connected vehicles, 6G smart terminals will deeply perceive their surrounding environment [13], enabling complex tasks such as intelligent obstacle avoidance and path optimization. To support these scenarios, the network needs to provide capabilities for sensing, collecting, and modeling physical world data, and offer high-quality data services for smart terminal interaction based on a unified data management framework.

2.3 Intelligent Collaboration

In scenarios like low-altitude economy and smart factories, the behavior patterns of 6G intelligent terminals are evolving from "individual intelligence" to "multi-intelligence collaboration" [14]. This involves collaborative tasks among multiple intelligent terminals, such as drone logistics delivery and robot collaborative handling [14]. To support these scenarios, the network needs to provide on-demand networking and dynamic task scheduling capabilities for multiple terminals, enabling fast response and collaborative decision-making for intelligent terminals, thereby enhancing the efficiency of complex collaborative tasks.

In summary, the new generation of intelligent terminal services for 6G imposes higher demands on network resource provisioning, data management, and collaboration mechanisms. There is an urgent need to optimize the 6G AI-native network architecture supporting edge-device collaboration and to research key technologies for 6G edge-device intelligent collaboration. Firstly, an AI-native intelligent network architecture that intrinsically supports data and model management needs to be built to enable key edge-device intelligent collaboration capabilities, providing network support for efficient and intelligent collaboration between 6G intelligent terminals and edge computing resources. Secondly, focusing on the data, model, and resource requirements of 6G intelligent terminal services, key edge-device intelligent collaboration technologies need to be proposed around the dimensions of data management, model collaboration, and heterogeneous resource fusion control.

3. Design of 6G AI-Native Network Architecture and System Evaluation Model

6G intelligent terminals possess AI-native, integrated sensing, and intelligent collaboration service characteristics, which necessitate an AI-native design for the edge-device collaborative 6G network architecture. To meet integrated sensing service demands, the 6G network must build a unified data processing framework that supports data collection, transmission, storage, and sharing, thereby providing on-demand data services for terminal, network decision-making, and edge-device intelligent collaboration. The 6G network must possess multi-level network global model control capabilities distributed across the edge and terminals, enhancing intelligent service performance metrics through multi-node collaboration. To satisfy diverse AI-native service demands, the 6G network must achieve integrated intelligent service scheduling and experience assurance for communication, computing, data, and models, supporting dynamic, on-demand edge-device heterogeneous resource allocation.

3.1 AI-Native Network Architecture

The current architecture, which separates mobile access and model management systems, cannot meet the demands of integrated sensing, intelligent and efficient collaboration, and multi-dimensional service on-demand scheduling in edge-device collaboration. This paper proposes an "AI-native" data-model-control network architecture. Through a unified distributed data-model management and heterogeneous resource fusion scheduling mechanism, it effectively supports the stringent performance requirements (e.g., ultra-low latency, ultra-high precision) of complex intelligent services for terminals.

Logically, the AI-native network architecture comprises the Infrastructure Layer, Model Management Layer, Resource Control Layer, and Service Orchestration Layer, supporting diverse intelligent applications through unified interfaces.

The Infrastructure Layer consists of distributed network components such as intelligent terminals, edge computing nodes, and wireless access networks. It enables dynamic cross-node collaboration through ubiquitous network access, providing fundamental resource support for intelligent service access and edge-device collaboration.

The Model Management Layer achieves unified data and model management for distributed edge-device networks. Its core functions include: 1) Building a unified data control framework to achieve on-demand collection, transmission, storage, and processing of heterogeneous device data; 2) Implementing full lifecycle management of edge models, including training, deployment, selection, monitoring, and updating; 3) Establishing a global model deployment view based on sensing terminal model capabilities.

The Resource Control Layer achieves unified modeling and collaborative control of multi-dimensional resources including communication, data, models, and computing. Through atomic capability abstraction and integrated control mechanisms, it supports end-to-end heterogeneous resource integration and unified scheduling.

The Service Orchestration Layer receives diverse application requests externally and utilizes built-in AI models to generate dynamic edge-device collaboration orchestration strategies, including "demand analysis—resource mapping—task scheduling," to ensure on-demand service experience and maintain efficient system operation.

3.2 System Evaluation Model

Current edge-device collaborative inference strategies primarily focus on optimizing single metrics such as inference latency, inference accuracy, and energy consumption. There is a lack of system-level evaluation metrics for multi-level distributed networks in edge-device collaboration, making it difficult to achieve optimal model control and management for long-term system utility. To address these issues, this section designs a system performance evaluation method that considers intelligent collaboration and multi-dimensional service metric assurance, overcoming the limitations of traditional single-dimension evaluations and providing a model benchmark for overall system utility optimization.

The edge-device collaborative system evaluation model covers multi-dimensional metrics such as basic resource overhead, model collaboration efficiency, and energy efficiency ratio. Basic resource overhead includes communication bandwidth, computing power, and storage resource overhead throughout the collaborative inference service lifecycle. Model collaboration efficiency needs to consider inference latency, accuracy, and model complexity. The energy efficiency ratio focuses on the energy consumption per task. By constructing a full-process energy consumption model that covers energy consumption during communication transmission, computing processing, and model updating, and by using methods like dynamic weighting and normalization to convert metrics with different dimensions into comparable numerical values, a comprehensive multi-dimensional system evaluation model is established. Furthermore, the system evaluation model is flexible, allowing for dynamic adjustment of weight coefficients based on different service scenario characteristics (e.g., real-time scenarios may prioritize latency and reliability, while offline scenarios may focus more on energy efficiency) to meet the system utility optimization goals of different scenarios.

4. 6G Edge-Device Intelligent Data Management and Control

4.1 Dataset Construction

Large-scale, high-quality data is a fundamental resource for model research and development. On one hand, edge-device collaborative model deployment requires adaptation to wireless environments, which are characterized by high time-variability and complex propagation mechanisms. Traditional methods struggle to meet the low-latency requirements of edge-device intelligent collaboration services in terms of channel parameter estimation accuracy and real-time performance, posing challenges for high-quality channel data collection in real wireless environments. On the other hand, terminal decision-making for intelligent connected vehicles, intelligent drones, and intelligent robots involves complex interactions in dynamic environments. Their adaptability and generalization capabilities in different environments depend on the diversity of environmental data, presenting challenges in collecting diverse, high-quality, real-environment data. Therefore, dataset construction for real channel environments and physical worlds is a key technology for enabling intelligent interactions for 6G smart terminals.

For channel dataset construction for intelligent terminals:

• First, establish a data collection hardware foundation based on large-scale antenna arrays and high-performance signal processing platforms.
• Second, conduct channel data collection across diverse scenarios and perform preprocessing such as noise reduction and format unification on the raw data.
• Finally, construct high-fidelity, scalable real-scenario channel models by combining terminal antenna patterns with coupled computation methods for real wireless channel parameters and wideband channel responses.

For environment dataset construction for intelligent terminals:

• First, establish a multi-modal data sensing platform, including sensing base stations and LiDAR, to integrate sensing data from different devices.
• Second, leverage the advantages of multi-source data, such as communication signal sensing for large-scale environmental outlines and LiDAR point cloud measurements for pose data, to improve the efficiency and resolution of indoor and outdoor map building using data synchronization and registration algorithms.

4.2 Data Management Framework

Edge-device intelligent collaboration services require cross-device data fusion and cross-domain data sharing. However, existing data collection frameworks primarily focus on core network functions and lack cross-domain data management mechanisms for distributed multi-level edge-device networks. Therefore, a full lifecycle management framework is needed to support data collection, storage, and sharing, ensuring efficient and reliable interaction for edge-device multi-node intelligent collaboration tasks.

Regarding data collection, establish unified data format standards and standardized collection processes to support standardized data collection from heterogeneous devices. Optimize collection transmission efficiency by selecting adaptive transmission protocols based on data characteristics and traffic demands. For data storage, design storage architectures and indexing mechanisms based on data types and real-time requirements to support efficient data retrieval by network function modules and cross-domain nodes. For data sharing, define standardized cross-module and cross-node data sharing interfaces, adopt service-oriented invocation mechanisms to trigger sharing processes, and ensure sharing efficiency through adaptive transmission protocols.

5. 6G Edge-Device Intelligent Model Management and Collaboration

5.1 Reference Model Library Construction

Based on data collection and modeling, there is an urgent need to build AI models for various edge-device intelligent collaboration scenarios. Currently, significant differences exist in technical solutions and model designs across different use cases and vendors, hindering the standardization and industrial deployment of intelligent collaboration among multiple edge-device nodes. Therefore, there is a pressing need to build a 6G+AI reference model library that provides benchmark and comparison models for typical scenarios. Through reference model testing and validation, it can guide the industry in achieving optimal trade-offs between model performance and complexity. First, build reference models for typical 6G edge-device intelligent collaboration use cases, providing reference model design routes and model architectures and parameters to ensure model reproducibility. Second, establish a model validation and evaluation simulation platform, integrating open-source datasets and evaluation toolchains, to achieve full-process testing and establish model evaluation standards for a unified evaluation system.

5.2 Edge-Device Model Collaboration

With the development of intelligent connected vehicles, robots, and other next-generation intelligent terminals, intelligent applications demand higher performance across multiple dimensions such as low latency, low overhead, and high precision inference. However, the capabilities of individual terminals are limited, making it difficult to support high-precision inference for large models. Furthermore, edge static model inference is constrained by dynamic wireless environments, failing to adapt to dynamic channels and diverse service demands. Therefore, there is an urgent need to build edge-device model collaboration mechanisms that optimize model deployment and dynamic model collaborative inference strategies by leveraging the distributed and differentiated capabilities of edge-device nodes.

For edge-device model deployment, considering the limited and distributed computing power of edge nodes and terminals, spatio-temporal driven distributed edge-device model deployment mechanisms need to be designed, taking into account factors like edge node distribution, terminal spatio-temporal distribution, and service request spatio-temporal distribution. For specific service inference scenarios, monitor the completion status of model collaboration metrics. When metrics decline, trigger data collection tasks to edge-device nodes on demand through the data management framework. Then, use reinforcement learning to fine-tune and update the fused data and the original deployed model, thereby enhancing the model's adaptability and generalization capabilities for diverse services.

For edge-device model collaborative inference, the focus is on designing network-assisted edge-device model collaboration mechanisms that enable efficient intelligent model collaboration inference through distributed multi-level computing technologies, supporting high computing power demands for intelligent services from terminals.

Specifically, design edge-side model splitting and inference task allocation methods adapted to dynamic channels. Combined with the model deployment status, optimize edge-device model selection strategies to meet multi-dimensional requirements such as inference accuracy, latency, and energy consumption for edge-device inference services.

6. 6G Heterogeneous Resource Fusion Control and Scheduling

6.1 Heterogeneous Resource Fusion Control and Scheduling

6G intelligent services have diverse demands for communication, computing, data, and models. Terminal distribution and service requests are dynamic and time-varying, requiring the network to provide on-demand resource and service scheduling to accurately respond to terminal service requests. Currently, edge computing service offloading is primarily based on service deployment and terminal location information, and model scheduling relies on centralized control by application service providers. This makes it difficult to achieve fused control and scheduling of heterogeneous resources such as communication, computing, data, and models, hindering efficient resource utilization. Therefore, there is an urgent need to build edge-device heterogeneous resource fusion control and scheduling mechanisms and design edge-device heterogeneous resource fusion control mechanisms adapted to dynamic service demands.

First, perceive the resource status of distributed computing nodes on the terminal and edge sides in real-time to generate a global resource status view. Second, upon receiving a service request, generate an optimal integrated strategy for data collection and transmission, edge model selection, and communication and computing resource allocation based on service demand and terminal mobility. This strategy aims to ensure service metrics while improving overall network resource utilization.

6.2 Task-Driven Flexible Networking

For multi-intelligent terminal collaborative tasks such as robot collaboration and intelligent connected vehicles, multiple intelligent terminals need to complete their individual intelligent services while jointly accomplishing multi-terminal collaborative tasks. For example, reference [15] optimizes model splitting, networking, and resource allocation strategies for vehicle-to-vehicle and vehicle-to-infrastructure collaboration in edge-device collaborative vehicular networks to reduce AI task latency. Therefore, for multi-intelligent terminal collaboration scenarios, it is necessary to accurately identify collaborative task requirements in terms of data, models, etc., dynamically form task-driven multi-terminal-multi-edge node collaborative networks, and allocate shared heterogeneous fused resources to achieve efficient interaction and collaboration among multiple intelligent terminals and between edge-device multi-nodes.

7. Edge-Device Intelligent Collaboration Empowers Multi-Intelligent Terminal Collaboration

In scenarios such as emergency rescue, urban security, and intelligent connected vehicles, multi-intelligent terminals need to collaborate to achieve data fusion and unified intelligent decision-making. Taking urban security as an example (illustrated in Figure 2), ground robots capture and identify ground targets, drones perform aerial environment perception and target identification, and smart glasses capture human viewpoint footage and perform identification. Each terminal performs single-view information collection and intelligent identification. However, the single-view perspective and limited computing power and battery capacity of individual terminals make it difficult to support the rapid search and global judgment demands of urban security. With 6G edge-device intelligent collaboration, the capabilities of these three terminals can be optimized, allowing edge nodes to share computing demands. This enables collaborative processing across multiple services, efficiently completing intelligent interconnection and fused decision-making tasks. Therefore, the edge-device collaborative 6G AI-native network is a crucial means to solve complex intelligent service challenges.

8. Conclusion

With the evolution of 6G intelligent terminals and intelligent services, edge-device intelligent collaboration has become a powerful means to support the complex AI service demands of 6G intelligent terminals. By building a unified edge-device data-model control framework and optimizing edge-device intelligent collaboration and heterogeneous resource fusion control mechanisms, 6G networks can effectively support the intelligent and collaborative services of multi-form terminals such as AI phones, intelligent robots, and intelligent connected vehicles, addressing challenges like low latency and high precision. In the future, with continuous technological evolution and the refinement of the industrial ecosystem, the deep integration of 6G and AI will further unlock the potential of intelligent terminals, providing stronger technological support for vertical industry applications and societal digital transformation, ushering in a new era of "Internet of Intelligence."

References