IoT Think Tank Original
On July 2, the 2026 Advantech Industrial AI Ecosystem Partner Summit was held in Suzhou under the theme of "Edge Intelligent Computing · Co-creating the Future." The summit gathered nearly 600 ecosystem partners to focus on AI implementation in scenarios such as smart manufacturing, smart energy, robotics, smart devices, and smart cities, through keynote speeches, two roundtable discussions, four industry sub-forums, and application exhibitions.
Rather than continuing to discuss model parameters, algorithm capabilities, and computing power scale, this summit more frequently addressed another set of questions: How can AI enter real devices and production processes? How can the closed loop from perception to execution be completed? And how can a POC (Proof of Concept) be replicated across more production lines and factories?
Behind these questions lies an overall shift in the focus of industrial AI.
Eight years ago, when Advantech held the IoT Co-creation Summit in Suzhou, the keyword was still the Industrial Internet, focusing on how to connect devices, systems, and data. Eight years later, the industrial proposition has further evolved from "connectivity" to "intelligence and autonomy." In his opening address, Luo Huancheng, General Manager of Advantech (China), made a core judgment: Industrial AI is undergoing a continuous evolution from machine learning and generative AI to AI Agents, and then to Physical AI. As AI gradually acquires autonomous perception, decision-making, and execution capabilities, the industry is moving from digital intelligence to the physical world, and industrial sites will enter a new stage of autonomous decision-making.
From Assisted Decision-Making to Autonomous Collaboration: Physical AI Enters the Industrial Frontline
Physical AI is not an overthrow of existing industrial AI, but an expansion of its capability boundaries. Traditional industrial AI mostly plays the role of identification, prediction, analysis, and recommendation; generative AI and large models have enhanced machines' understanding of knowledge, language, and complex information; AI Agents further possess the capabilities of task decomposition, tool invocation, and process scheduling. When these capabilities are connected with sensors, robots, controllers, and industrial software, AI begins to move from the digital world into the physical world.
In other words, the key to Physical AI is not just to make machines "smarter," but to make them "understand the site better"—understanding equipment status, process flows, environmental changes, and safety constraints, and translating judgments in the digital world into actions in the physical world.
Robots are the most intuitive carriers of this trend. Thirty years ago, the core goal of robotic arms was to replace repetitive labor and improve efficiency; today, AMRs (Autonomous Mobile Robots), AGVs (Automated Guided Vehicles), collaborative robots, and even humanoid robots need to simultaneously process multimodal information such as vision, voice, radar, navigation, and motion control, and complete planning, obstacle avoidance, and task execution in dynamic environments. From applications in robotics, medical imaging, smart transportation, and energy management, Physical AI is no longer just a concept for the future, but is gradually entering industrial practice supported by Edge AI capabilities. It is more like the result of industrial AI moving from assisted decision-making to autonomous collaboration.
However, to truly link perception, reasoning, and execution into a closed loop, models alone are not enough. AI needs to understand tasks, invoke tools, and coordinate multiple systems, which is exactly the role AI Agents play on the industrial frontline.
AI Agents Enter the Factory: From Point Tools to Process Collaboration
If Physical AI outlines the direction for industrial AI to enter the physical world, then AI Agents are the collaborative hub connecting tasks, tools, and processes.
In consumer and office scenarios, Agents are usually understood as assistants capable of conversing, searching for information, or completing tasks on behalf of users. However, on the industrial frontline, Agents face equipment, systems, processes, and organizational workflows. Their core task is not to "answer questions," but to understand goals, decompose tasks, invoke tools, coordinate multiple systems, and continuously adjust based on on-site feedback.
Advantech's answer is the "AI Factory Brain," which depicts a possible path for multi-agent collaboration to enter factory operation processes: by linking tasks among different specialized Agents, it drives the factory from rule-driven traditional automation further toward an adaptive system with continuous optimization capabilities. For example, when a quality Agent detects an anomaly, it not only generates a report but can also link with an engineering Agent to analyze process causes, notify a scheduling Agent to determine whether to adjust the production plan, and then have an energy Agent re-optimize the load based on capacity changes. Data and actions originally scattered in various systems are coordinated by multiple specialized Agents to form a cross-system business closed loop. Its value lies less in replacing a specific position and more in connecting originally siloed data, systems, and departments.
Therefore, the true industrial value of AI Agents lies not in creating more dazzling demos, but in precipitating industry experience into reusable skills, software components, and standard products, enabling intelligent capabilities to enter real processes and be continuously delivered.
For industrial Agents, the ability to answer questions is just the starting point; the real test is whether they possess clear permission boundaries, credible data basis, and auditable execution processes. Whether the constraints of these permissions, data, and executions can be completed at the millisecond level depends on which layer the intelligence runs on—which brings the issue to the edge.
Edge AI Supports the Frontline: Intelligence Needs to Be Closer to Where Data is Generated
For Physical AI to enter the industrial frontline, it cannot rely solely on cloud-based large models.
Industrial scenarios have a series of "hard constraints": robots and production lines require low-latency responses, production data involves security and privacy, network environments are not always stable, and equipment must run continuously for long periods. Meanwhile, the site has a large number of cameras, sensors, PLCs (Programmable Logic Controllers), controllers, MES (Manufacturing Execution Systems), WMS (Warehouse Management Systems), and SCADA (Supervisory Control and Data Acquisition) systems, with highly heterogeneous protocols, chips, and data structures. If all data is uploaded to the cloud and then waits for a return, many scenarios will struggle to meet real-time, stability, and security requirements.
Therefore, wherever data is generated, it is even more necessary for intelligence to be close to where it occurs. The significance of Edge AI is not just moving computing power from the cloud to the site, but enabling devices to complete real-time perception, reasoning, and feedback locally, ensuring that sensitive data does not all have to leave the factory, and allowing systems to maintain basic operations in weak-network or even offline environments.
Centered around this demand, Advantech showcased a multi-layered hardware architecture ranging from Edge AI acceleration modules and Edge AI all-in-one machines to Edge AI Servers at the summit.
Supporting this hardware is the WEDA (WISE-Edge Developer Architecture) edge intelligence developer architecture. As a containerized development and runtime environment, WEDA helps developers lower the threshold for Edge AI development and deployment through unified interfaces, MCP (Model Context Protocol), and AI Skills, enabling more efficient deployment of models into industrial scenarios such as manufacturing, energy, healthcare, and cities.
The on-site roundtable also outlined a "simulation-training-inference" closed loop: first using digital twins and virtual environments to generate and validate scenarios, then using training systems to optimize models, and finally deploying the models to edge devices and robots for operation; new data generated from the real site then flows back into the simulation and training stages, forming continuous iteration. This mechanism of "from virtual to reality, and then reality feeding back into virtual" can significantly reduce the trial-and-error and R&D costs of physical equipment.
Cloud-based large models excel in general knowledge and complex reasoning, while the edge handles real-time response, device connectivity, and execution determinism. The two are not a substitute relationship, but rather cloud-edge collaboration.
If cloud-edge collaboration solves the problem of how AI enters the frontline, then the next harder step is how to stably replicate the same set of capabilities across different devices, production lines, and factories.
From POC to Scale: The True Threshold of Industrial AI Lies in Engineering
For industrial AI, creating a demo is not difficult; the hard part is replicating it across a hundred production lines.
The POC stage often has clear boundaries: data comes from a small number of devices, the environment is relatively fixed, and the system can rely on manual maintenance. However, once it enters scaled deployment, cameras, sensors, chips, SDKs, interface protocols, lighting conditions, materials, and production takt times may all change, and the same solution may require re-adaptation when moved to a different factory.
Industrial AI has long faced the problem of "effective at a single point but difficult to replicate." The root cause is not necessarily insufficient model accuracy, but rather that device access, data governance, system integration, industry knowledge, and post-operation maintenance have not yet formed standardized capabilities. It is precisely for this reason that the true watershed for the scaling of industrial AI is often not just whether model accuracy can be improved from 98% to 99%, but whether the solution can be stably replicated across devices, factories, and regions.
Centered precisely on the proposition of replicability, Advantech emphasizes the path of platformization and ecosystem collaboration. Advantech provides hardware, development platforms, and foundational software, while algorithm companies, system integrators, and industry partners supplement professional knowledge, skills, and solutions. Advantech Chairman Liu Kezhen cited the AI "five-layer cake" framework on-site, dividing the industrial chain into energy, chips, infrastructure, models, and applications, and explicitly stated that Advantech hopes to join forces with partners to focus on developing the industrial AI application layer. The cases of food manufacturing, electronics manufacturing, energy management, and smart spaces at the summit also illustrate that truly valuable solutions come from the long-term co-creation of platform enterprises, industry partners, and end-users.
In the AI era, the industrial ecosystem is no longer just a sales channel, but a co-production mechanism for technology, data, industry knowledge, and scenario experience. Physical AI involves chips, computing hardware, sensors, industrial software, models, system integration, and end-users; it is difficult for any single enterprise to independently complete the entire chain.
Conclusion: The Competition in Industrial AI is Shifting Towards Engineering and Scaled Replication
Looking at this summit as a whole, three increasingly clear industrial judgments can be seen:
Physical AI is not an overthrow of existing industrial AI, but a further addition of understanding, decision-making, and execution capabilities on top of machine vision, equipment operation and maintenance, digital twins, and industrial software;
AI Agents and Edge AI are indispensable; the former undertakes task understanding and process coordination, while the latter is responsible for real-time perception, local reasoning, and device connectivity;
And the focus of competition in the next stage of industrial AI will shift from single-point technical capabilities to engineering and scaled replication.
As Advantech Chairman Liu Kezhen said, "Entering the AI era, the focus of industrial development is gradually shifting towards platform capabilities and ecosystem collaboration. Advantech will continue to join hands with global partners to jointly promote the scaled implementation of industrial AI in more industrial scenarios."
The discussion on industrial AI is moving away from a mere competition of model parameters. What the industry truly needs is not just a model capable of generating content or identifying anomalies, but a set of system capabilities that can enter devices, integrate into processes, accept constraints, and run stably across different sites.
From this perspective, Physical AI outlines the direction, Edge AI supports the frontline, AI Agents connect the processes, and platformization and ecosystem collaboration determine whether these capabilities can ultimately move towards scaled application.