At the 2026 World Robot Conference (WRC), held from August 19-23 at the Beijing Yichuang International Convention and Exhibition Center, a special session on "New Paradigms of AI Large Models Empowering Robotics and Embodied Intelligence Industries" was held alongside the main event. Liu Tianwei, Vice President and Marketing Director of Huiwen Robotics, attended and delivered a keynote speech, sharing his insights into the company's practices in applying large models to industrial environments.

Liu Tianwei began by noting that large models have developed rapidly over the past two years, with robots evolving from basic perception, navigation, and control toward understanding tasks, making decisions, and learning skills. However, he emphasized that regardless of a model's capabilities, it ultimately must be integrated into a real robot and a real factory to complete actual tasks. He expressed his desire to share how Huiwen Robotics has already entered factories, transforming large models into components of industrial agents.

Before discussing large models, Liu raised a fundamental question: why does a customer buy a robot? It is not because of larger model parameters. Customers evaluate practical aspects: whether tasks can be completed reliably, not just in demonstrations; whether the robot can operate stably over the long term; whether it can recover from anomalies; and whether it can coordinate with existing production systems such as MES and WMS. Ultimately, they assess whether the solution can be deployed, delivered, maintained, and supported after-sale. Internally, Huiwen consistently believes that technical capabilities must evolve into a robot product that customers are willing to use long-term, which serves as the starting point for their understanding of industrial agents.

Huiwen's approach to industrial agents does not involve building a new industrial robot from scratch. Instead, it builds upon their existing AMR product line already deployed in industrial settings. Liu introduced their "1-N-X" system architecture: at the base lies a shared set of robotic technical capabilities, above which sit different robots with varying payloads and positioning, such as the X100, X300, X600, and AMR forklifts. These forms can adapt to different tasks—whether flatbed, lifting, or multi-layer pallet handling. Further up, these robots must integrate into customers' production systems, connecting to MES, WMS, ERP, and other scheduling platforms. Therefore, Huiwen views robots not merely as tools to complete isolated tasks, but as entities that must shoulder real responsibilities once integrated into production systems.

Breaking down an AMR system, it comprises a complex software and hardware stack: sensors, drives, power supplies, and communication devices at the bottom, followed by mapping, localization, path planning, environmental perception, obstacle avoidance, and system integration. Above that lie task distribution, multi-robot scheduling, traffic control, and other functions. Features like rapid deployment, SLAM navigation, local ad-hoc communication networks, and cloud-based scheduling are all products of a long-accumulated engineering system. Liu stressed that integrating large models into robots is not about adding intelligence to a blank canvas—there is already a mature and complex robotic engineering system in place.

Once robots enter actual factories, a key observation emerges: while a screen may simply show a robot moving from point A to point B, customers care about far more than that movement. A single delivery task, for example, might originate from an MES, WMS, or manual call, then enter a scheduling system for assignment to an AMR. After navigating and transporting, the robot must hand off to racks, pallets, or production equipment, and finally feed back task status, robot status, and equipment status. The site may also include automatic doors, elevators, access controls, and other devices. A seemingly simple transport task thus involves a complete production task chain, where robots execute not isolated actions but specific tasks within a production system.

As robots become more deeply integrated into production, the complexity does not increase linearly. With a single robot, the focus is on localization, navigation, obstacle avoidance, and task execution. With multiple robots, issues arise around task prioritization, traffic conflicts, and congestion, requiring multi-robot scheduling and coordination. Once truly embedded in production systems, robots must also handle workstations, equipment, WMS integrations, rush orders, and anomaly handling. Liu noted that after years of AMR work, a key realization is that the deeper robots go into production, the more the competitive focus shifts from individual robot performance to the entire system's long-term stable operation. This provides a realistic boundary for where large models can add value in robotics.

Today's industrial robots are highly capable—they can map, localize, plan paths, avoid obstacles, execute tasks, and even operate in distributed ad-hoc networks of a dozen or more units in weak or no network environments. However, this relies on a critical premise: tasks are already defined, with robots knowing where to go and what to do, and most rules and boundaries are pre-configured. Current industrial robots excel at reliably executing clearly defined tasks. The next question is: what happens when tasks are not fully defined? When environments contain constantly changing areas? If a robot is only told the desired outcome, can it independently determine next steps? Liu believes this is where large models begin to create real value.

Real factories contain many elements that cannot be pre-scripted. Tasks change, environments shift, and rush orders alter targets, priorities, and resource constraints. More complex tasks may require robots to combine multiple skills. At this point, robots need more than execution—they need to understand, reason, decompose tasks, plan, select skills, and dynamically adjust based on real-time changes. Huiwen's understanding is that large models are not for rewriting or optimizing existing rules, but for handling areas where all possible variations cannot be exhaustively enumerated in advance.

A crucial question is where large models should integrate into the robot stack. Liu simplified the stack into layers: at the top are industrial tasks—what ultimately needs to be accomplished. Below that is task understanding and decision-making—comprehending goals, understanding scenarios, decomposing tasks, and planning steps. He believes the greatest incremental value of large models occurs at this layer. Further down are robot skills—navigation, movement, obstacle avoidance, and docking—which robots can call upon. At the bottom are real-time control and the robot body itself. Technical approaches like VLA are valuable for connecting upper-level understanding and decision-making with lower-level skills and actions. However, feedback is crucial in industrial scenarios: after each step, the robot must assess whether the environment changed, whether the task succeeded, and whether the next step remains valid. Thus, large models entering robots do not replace the entire technical stack; they add higher-level understanding and decision-making capabilities on top of the existing robotic engineering system.

Transitioning from large models to true industrial agents requires crossing several stages: first, understanding tasks—for example, knowing that materials must be transported to a specific workstation, what the goal is, and what constraints exist. Second, planning tasks—breaking a goal into executable steps and re-planning as conditions change. Third, invoking relevant skills. Finally, physical execution. Execution is not the endpoint; robots must also judge whether the step succeeded, whether conditions changed, and whether the original plan remains valid. Huiwen believes the real key to industrial agents is not merely understanding a spoken command, but consistently translating understanding and decision-making into tangible action results in the real world.

Another essential capability is continuous learning. Completing a task does not mean the state or problem is permanently solved. Real factories generate new environments and anomalies daily. The operational process itself produces valuable data—what tasks ran, what site conditions existed, and how robots executed. This data must re-enter model and skill training, then return to real robots for validation. Huiwen aims to create a cycle where real tasks generate real data, real data drives intelligent evolution, and stronger intelligence returns to real tasks. The focus is not just data volume, but whether the data genuinely helps robots perform better on subsequent tasks.

This explains why Huiwen is pursuing industrial agents: not to overturn previous work and build new robots, but to build upon existing capabilities that remain essential—robot bodies, perception, navigation, motion control, multi-robot coordination, and system integration. These determine whether robots can truly enter factories. With large models, Huiwen aims to add task understanding, scenario comprehension, multi-stage decision-making, complex skill combination, and capabilities for VLA and data learning. From their perspective, industrial agents are built on mature, highly available robotic capabilities, with intelligence growing upward. The more solid the foundation, the more likely intelligence can penetrate real industrial sites.

Factories present diverse tasks that cannot be solved by a single robot type. AMRs handle small parts, intelligent forklifts manage pallet transport, heavy-duty robots tackle heavy loads, and wheeled dual-arm robots address picking and sorting. Liu's judgment is that tasks determine the form factor—no single robot configuration solves all industrial tasks. However, while bodies may differ, intelligence capabilities should be gradually shared: how to understand tasks, decompose goals, plan steps, and adjust to changes. "One brain, multiple machines" does not simply mean using one large model to control different robots; it means developing a system of task understanding, planning, decision-making, skill execution, and data learning that can be reused across different robot platforms.

Returning to productization, Liu outlined Huiwen's chosen path. The first step is product deployment: robots must be real, mature products that can be deployed, connect to workstations and systems, and genuinely enter factories. The second step is task execution: not a one-time demo, but long-term, stable assumption of production tasks. The third step is data accumulation: as robots operate continuously, they generate diverse task states, environmental conditions, execution results, and anomaly data. The final step is intelligent evolution: this data—especially real-world data—drives improvements in task understanding, decision-making, skill learning, and VLA training and validation, with stronger intelligence returning to the factory. This forms a closed loop: product deployment, task execution, data accumulation, and intelligent evolution, returning to more real tasks. This reflects a clear productization philosophy: not building the largest possible model first and then finding uses, but letting intelligence grow layer by layer within real products and real tasks.

Liu concluded with three words summarizing Huiwen's understanding of large models entering robotics. First, "ROBOT": there must be a mature, real, and reliable robot product. Second, "TASK": robots must genuinely enter production flows—transport, delivery, turnover, and manufacturing—and bear real industrial tasks. Third, "INTELLIGENCE": robots must continuously improve their understanding, planning, decision-making, and skill capabilities within real tasks, iterating based on feedback from results. Only by truly combining these three elements can what they call an industrial agent emerge. The endpoint of large models entering robotics is not a more polished or clever demo, but an industrial product that can enter real factory sites, continuously complete work, and evolve over time. This is what Huiwen Robotics has done, is doing, and will continue to do.