China Unicom's Innovations and Practices in Industrial Intelligence

In recent years, next-generation technologies like AI have accelerated the intelligent transformation of the industrial ICT sector. In China’s 15th Five-Year Plan, policies such as AI + industrial Internet and AI + manufacturing are set to propel AI deeper into the heart of industrial production. As both a builder of China's information infrastructure and an enabler of industrial development, China Unicom has analyzed the core challenges in intelligent transformation, including data processing, task execution, and robotic decision-making. Based on these insights, China Unicom has developed a full-stack solution that spans data governance, agent networks, and embodied AI.

The three phases of China Unicom's industrial intelligence

China Unicom's industrial journey unfolded in three phases: first, connecting everything with 5G; then, sensing through data and vision; and now, cognizing with AI and agents.

Since 2018, the large-scale 5G technology adoption has been fundamentally changing networking. Through its 5G private networks, China Unicom has delivered low-latency, high-bandwidth connectivity to virtually every aspect of industrial operations. Production lines are now capable of advanced visual sensing and identification. Mobile terminals have freed production lines from traditional constraints. Autonomous guided vehicles (AGVs), AR glasses that merge the virtual and real worlds, and centrally controlled workstations can be flexibly reconfigured to form new flexible production lines. To date, over 9,500 5G-powered production facilities have been established in China, spanning industries from manufacturing and mining to ports and power plants. These smart facilities are also extending to other markets, like Thailand (Figure 1).

Figure 1: Midea's 5G factory in Chon Buri, Thailand

In 2022, the focus of digital transformation shifted to data-driven development. China Unicom launched the Unilink Industrial Internet Platform, centered on IoT and big data, to carry the device and management data of industrial enterprises. This platform offers protocol parsing and structured modeling services, which enhances system interoperability by allowing upper-layer applications to access data securely without having to navigate the complexities of underlying devices and systems. Unilink has been included in the Ministry of Industry and Information Technology's (MIIT's) Cross-industry and Cross-domain Industrial Internet Platform List for three years. It rose from 29th in the rankings to 9th in 2024, and entered the top 6 in 2025.

In 2024, the AI boom really started to take off. What truly drives users to choose AI is not parameters or compute power, but how seamlessly it integrates into existing production scenarios. The forms of intelligence are evolving into vertical foundation models that can understand professional data and generate accurate answers, agents that automate tasks, and embodied AI robots that enable human-machine collaboration. Building an open industrial intelligence ecosystem on top of its networking and digitization foundation is the goal of the Unilink platform evolution.

The three core pain points of industrial intelligence

Despite its promising prospects, industrial intelligence still faces three major challenges in practical implementation in factories, mines, and ports.

The first challenge is the semantic gap in data. The foundation model needs to understand a vast amount of proprietary, highly specialized enterprise data and ensure that generated content strictly adheres to industry specifications, safety standards, and process requirements. Otherwise, intelligence would remain at the Q&A level and be unable to guide decision-making.

The second challenge is task breakdown and execution. Conversational capabilities alone are insufficient. The foundation model needs to break down user intentions into a series of executable tasks, and the operating system needs to invoke different agents and enterprise applications to complete these tasks automatically.

The final challenge is the lack of generalization in robots. Traditional robots lack perception, while new humanoid and quadrupedal robots have insufficient generalization. They lack the autonomous reasoning and decision-making needed for industrial settings. The key to full-scale intelligence is enabling embodied intelligent robots to not only perform tasks, but work in close collaboration with humans and other automated systems.

China Unicom details three solutions for industrial intelligence

To tackle these challenges, China Unicom has detailed three technical approaches, based on data, collaboration, and terminals.

First, an AI-ready data governance system: Enterprise data is multimodal. It includes device status data (temperature, pressure, current, rotational speed, etc.), production data (customer, order, product, material information, etc.), diverse system-generated data (file, program code, API, process, etc.), and unstructured data (image, audio, video, paper document, etc.). China Unicom has introduced its own data ontology to enable foundation models and agents to understand and reference this data.

Instead of simply collecting data, China Unicom uses semantic modeling and knowledge graphs to extract and transform multimodal data into semantic information and knowledge. Working with industry customers, China Unicom aligns the semantics of data in specific settings to ensure that relationships across different dimensions are unified in a single vector space. This approach enables foundation models to make informed decisions grounded in industry common sense, not just probabilistic guesses.

China Unicom's equipment team partnered with tech companies to create a specialized model for the production of China's C919 large aircraft. This project demonstrated significantly improved decision-making precision for advanced manufacturing. In the fashion sector, China Unicom developed a data-driven fabric performance predictor for Shanghai Challenge Textile that greatly reduced the time to market for new materials.

Second, enterprise-level Internet of agents (IoA): The number of job-specific and task-specific agents is growing rapidly. Traditional IP address-centric and static rule-based security models are insufficient for dynamic agent migration, autonomous decision-making, and efficient collaboration. To tackle this issue, China Unicom has proposed the enterprise-ready IoA, a technology system akin to the traditional Internet and intranet but fully AI-native and independent.

First, addressing is evolving from traditional location-based addressing to identity-based addressing. In conventional networks, servers are located using IP addresses. In the IoA, system routing matches agents directly by their identities, types, and permissions, eliminating reliance on destination IP addresses.

Second, security is transitioning from perimeter protection to intrinsic security. Traditional models are like fortresses, mainly defending against external threats. In the IoA, however, each agent can become an entry point or target for attacks, making a comprehensive security mechanism covering the entire network essential. This mechanism must include:

• Continuous identity-based authentication and dynamic authorization for each agent interaction request.
• Dynamic adjustment of access permissions for agents based on their identities and real-time risk assessments (such as behavioral anomalies and environmental changes), enforcing the principle of least privilege to prevent permission abuse.
• A full-link verifiable trust system covering the hardware execution environment (such as TEE), code integrity, and identity credentials. This establishes a trusted data space and ensures that agents' actions match their claims and are trustworthy.
• Recording key agent operations, decision-making chains, and communication content in a tamper-proof audit ledger, forming a complete evidence chain for backtracking and accountability.

At the Mobile World Congress (MWC) 2026, China Unicom launched UniClaw, an AI-native communication service and the latest addition to the Yuanjing Wanwu intelligent agent platform. It transforms traditional communications like phone calls and SMS into entry points for executing agent commands (Figure 2). At a March 2026 event in Huangpu District, Guangzhou, China Unicom presented its full-stack AI + New Industrialization Solution. This solution's ClawdSecbot security system mitigates risks such as permission miscontrol and plugin poisoning in OpenClaw deployment, ensuring secure AI for manufacturing enterprises.

Figure 2: China Unicom launching UniClaw

The original IP address-based routing network will be upgraded to an intelligent network for agent interconnection based on intent understanding, built-in security authentication, and compute-network collaboration. This will enable us to build an open ecosystem of industrial intelligence on existing network and digital foundations.

This transformation means compute-network collaboration is no longer just about moving computing power closer to data, but evolving into a super neural network that understands intent, autonomously schedules agents, and comes with built-in security features.

Building an open industrial intelligence ecosystem is the ultimate goal of compute-network collaboration in practice. At China Unicom Shandong, for example, China Unicom and Huawei have successfully deployed the autonomous network change and security agent. Using foundation model deep learning, it enables online simulation and risk assessment of network configuration changes. This transforms complex tasks that previously required manual checks into automated closed-loop agent operations, providing a solid underlying foundation for the open industrial intelligence ecosystem.

Third, an embodied intelligent brain operating across terminals: Previously, the Unilink platform enabled interoperability between applications and devices through protocol parsing and device models. Now, China Unicom has upgraded the platform to enable seamless interactions between robots (both traditional and embodied AI) and the embedded intelligent brain. This upgraded platform offers end-to-end services for embodied AI.

First, the platform accurately captures motion data. This involves deploying optical, inertial, and visual sensors to record the actions of human experts in fields like welding, material handling, and quality inspection. In the subsequent data adaptation and format conversion processes, the system uses dedicated algorithms to adapt captured joint space data to robot description files (such as URDF), bridging the gap from how humans do it to how robots do it.

In the model training phase, training samples are expanded to address the long-tail problem. Synthetic data of generalized actions is generated from real motion capture data and digital twin technology. A vision-language-action (VLA) embodied AI model is trained on this data to equip robots with visual sensing, language understanding, and decision-making abilities. In the virtual-to-real debugging phase, the trained model is deployed on robots using a high-bandwidth, low-latency 5G network. The robots achieve autonomous operations through reinforcement learning and fine-tuning from simulation to reality (Figure 3). The platform provides comprehensive technical support for robots using 5G network connectivity, sensory abilities (visual and auditory), and customized small model services.

Figure 3: CCTV, China's leading broadcaster, reports on the virtual-to-real training field of China Unicom's industrial embodied AI robots

China Unicom's exploration and practices in industrial intelligence spans from 5G private networks to smart factories, from industrial Internet platforms to edge-AI integration in core production — each step deepening our industrial intelligence capabilities. Going forward, we will strengthen our digital infrastructure; deepen the application of AI + manufacturing; and collaborate with industry partners across the value chain — including manufacturers, equipment suppliers, and model providers — to build an open, collaborative, and secure industrial intelligence ecosystem. This will transform industrial intelligence from isolated pilots to coordinated, widespread adoption and contribute to the growth of the manufacturing sector.