World Robot Conference (WRC) 2026 opened in Beijing on August 20 as China’s robotics sector entered a more demanding phase. The event coincided with Unitree Robotics’ IPO, which pushed the company’s market capitalization above RMB 400 billion (USD 59.4 billion) at its peak, adding momentum to an already crowded market.
This year’s WRC drew more than 300 exhibitors, up about 40% from a year earlier, with roughly 3,000 robots on display.
The newcomers included not only another fresh crop of startups, but also potential customers. According to 36Kr, 49 centrally administered state-owned enterprises attended as a group, bringing 12 real-world engineering scenarios across power grids, aerospace, mining, metallurgy, and other fields in search of robotics solutions.
WRC expanded to four halls, one more than last year. Hall C centered on complete humanoid systems, while halls A and D focused on deployment scenarios. Hall B featured upstream suppliers, together forming a rough cross-section of the embodied intelligence value chain. Global chipmakers including AMD, Infineon, and Realtek also appeared in hall B, seeking business from a new generation of robotics companies.
The conference offered a useful view of where the industry is heading. After walking the 52,000-square-meter venue, 36Kr found that exhibitors appeared to be under less pressure to sell a distant vision and more pressure to answer immediate questions:
- Does the model form a usable closed loop between data, decision-making, and execution?
- Can the robot perform real work?
With Unitree listed and several companies now valued above RMB 10 billion (USD 1.5 billion), embodied intelligence is entering a more exacting phase. WRC 2026 pointed to a renewed focus on commercial fundamentals. Whatever the long-term potential, companies still need to generate revenue in the present.
Moving toward deployment
At earlier WRC events, companies often competed over how “smart” their models were or how large their parameter counts had become. This year, the discussion was more practical. A recurring question was how to connect models, data, and physical tasks in a closed loop that improves performance during real-world deployment.
One technical shift was particularly visible. Last year, vision-language-action (VLA) models were the dominant theme. This year, exhibitors talked less about VLA itself and more about adding world models, systems designed to represent and predict how an environment may change, to VLA architectures.
A standalone VLA model can perceive a scene, interpret instructions, and generate actions, which can be sufficient for short tasks in controlled environments. But it does not necessarily model how its surroundings will change or how causal relationships unfold over time. As tasks become longer, errors can accumulate and performance can deteriorate in unfamiliar settings. World models are intended to address that weakness by helping a robot predict possible outcomes before acting.
Companies differ, however, on how VLA and world model capabilities should be combined.
One approach is to integrate predictive world modeling directly into the robot’s policy, or decision-making system. Under this architecture, action generation is informed by predictions about how the environment will change rather than functioning as a simple end-to-end mapping from input to action.
Yu Chao, founder of Lumos Robotics, told 36Kr that the company’s Prime R0 embodied control system is designed around that idea. It integrates a world model’s predictive capacity into the VLA decision framework so that semantic understanding, action execution, and situational forecasting operate in a single loop.
“We don’t treat the world model as a simple add-on,” Yu said. “We want it to learn not only the action, but also how the world changes after that action.”
Lumos also introduced NexCore, a robot skill evolution engine built around a similar cycle. It begins with real industrial tasks, collects real-world data, trains models to develop robot skills, deploys those skills back onto robots, and feeds new operational data and feedback into subsequent rounds of training.
Astribot is taking a different approach. Its model architecture emphasizes reasoning that can be interpreted and inspected, evolving from the explicit textual reasoning of Lumo-1 to implicit modeling of world dynamics in Lumo-2.
The tradeoff is largely between visibility and efficiency. In Lumo-1, the reasoning process is readable: the system organizes its reasoning in language before acting. Planning, prediction, and control are divided among three modules, with the world model serving as an intermediate predictive layer. That architecture can improve stability in long-horizon tasks, but moving data among modules slows the system and makes high-frequency interaction with the physical world more difficult.
With Lumo-2, Astribot shifted to what it calls a world action model framework. The model predicts changes in the world in latent space, an internal mathematical representation, and generates actions directly. The approach is faster but less transparent than text-based reasoning. The change reportedly makes inference 2.71 times faster than a standard autoregressive approach.
Whatever the architecture, the commercial question is the same: how does a model create value in a real operating environment?
One of the clearest changes at this year’s WRC was speed. Robots were visibly moving faster than in previous years, and several exhibitors attributed the improvement to extensive engineering optimization rather than any single breakthrough model.
Interfacing has been a recurring problem in earlier deployments. Model developers buying different robot bodies, or hardware companies procuring external models, have often had to build bespoke interfaces for each combination. The gap between hardware and software has remained wide.
Fourier Intelligence highlighted a framework at WRC that connects large language models, embodied models, navigation and planning systems, and low-level motion control. The aim is to let model developers plug in their systems for validation without rebuilding the full adaptation layer from scratch, effectively creating a common control stack.
Xu Shichao, technical director at Sunseed, said the push for faster robots also reflects a broader change in how companies use large models. In industrial settings, the sector is moving away from relying on a large model to deliberate over every step. Such an approach can make robots too slow while consuming excessive computing resources.
An emerging approach is a fast-slow control architecture. A larger model handles more complex reasoning, while smaller models manage high-frequency, repetitive actions. The division is intended to balance computing requirements, cost, thermal constraints, and response time without invoking the largest model for every movement.
Xu said companies are also increasingly using physical-world simulation and anticipatory inference so robots do not have to calculate every decision while already in motion. The goal is to raise throughput enough for production environments such as food processing, where speed directly affects operating economics.
Zhang Xiaoming, executive vice president of Coowa, pointed to another characteristic of embodied models. He said that while large language models are typically trained in a laboratory before being deployed at scale, robot models do not necessarily have to be fully trained before entering real environments. They can begin operating, accumulate data during work, and use those data to improve their models.
Coowa said more than 10,000 of its robots are in routine operation across over 50 cities and regions worldwide, covering multiple urban scenarios and generating more than 50 petabytes of real-world data. The company says it uses those data to improve model decision-making and generalization.
“Model improvement and scenario deployment happen in parallel, not one after the other,” Zhang said.
The faster robots at this year’s WRC therefore do not necessarily indicate a breakthrough at the model level. Instead, they may reflect greater emphasis on closing the loop between reasoning and execution, as well as engineering around the trade-off between slower high-level inference and faster physical action.
The limitations remain significant. Jiao Jichao, vice president of UBTech Robotics and head of its research institute, told 36Kr that the industry still lacks sufficient volumes of real-world interactive data and that high-quality data remain scarce. There is also no consensus on which types of data embodied models need or what training standards should look like.
Some teams are responding by testing not simply whether a model can complete a task, but where its capability boundary lies and at what level of difficulty it begins to fail.
Muka Robotics, for example, used three interactive demonstrations to probe different limits: curling, a cup-and-ball guessing game, and a maze. Curling tests how quickly a world model adapts to a dynamic environment. The guessing game tests multi-path simulation by moving a hidden ball among three cups and asking the robot to reason through possible future states before choosing an action. The maze is a longer-horizon stress test that requires consistent visual reasoning as the path becomes more complex, without allowing errors to accumulate.
The need for such structured games also illustrates how early embodied models remain. Even testing their capabilities in uncontrolled environments is difficult.
“The direct result of stronger models is that some scenarios that could not be delivered with rule-based systems and planning can now be delivered,” Fu Qiang, partner and head of product at Galbot, told 36Kr. “But there is still a gap between what’s deliverable, and what can be delivered at scale.”
For nontechnical observers trying to determine whether a robot is genuinely model-driven rather than preprogrammed, one model vendor suggested three checks:
- First, look for staff operating a remote controller or computer behind the demonstration.
- Second, see whether the robot simply moves between fixed points or repeats an identical task.
- Third, move an object unexpectedly and observe whether the robot continues toward the original location or stops and replans.
Home robots multiply as industrial deployments advance
The second major shift across WRC’s four halls was the proliferation of household scenarios.
Companies demonstrating household uses included Spirit AI, RobotEra, Astribot, Morphi Robot, Haier Robotics, Yunji Technology, Fourier Intelligence, Futuring Robot, X Square Robot, Galbot, and GigaAI.
Rather than presenting robots mainly as companions, exhibitors built living rooms and laundry rooms on the show floor and put robots to work.
Spirit AI, for example, demonstrated a living room tidying function. A user could give a general instruction and the robot would break it into subtasks: place a bowl in the dishwasher, return a can of cola to its place, put a toy in a bin, discard trash, map the environment, sequence the tasks, and complete them one by one.
The expansion of household demonstrations is being driven in part by world models and improving data availability. A Morphi Robot staff member told 36Kr that world models help robots predict future changes, while wider use of egocentric data collection, in which data is captured from the robot’s or operator’s point of view, has made everyday household data cheaper and more abundant. The staff member said an increasingly common rule of thumb is to pretrain models on tens of thousands of hours of data before fine-tuning them for specific tasks.
Beyond the technology, home deployment matters because many industry participants see it as a test of whether embodied intelligence can develop into a mass market. One exhibitor argued that the sector can be sustained over the long term only if robots eventually enter ordinary homes and personal robot ownership becomes commercially viable.
For now, however, nearly every household setup at WRC remained a demonstration rather than a deployment in an ordinary residence. Tasks often took 10–15 minutes, and some robots could not operate continuously because of battery or computing constraints, according to 36Kr’s observations. On the afternoon of August 20, heavy network use around the concentrated household demonstrations in hall C also constrained bandwidth and prevented some demonstrations from running.
A Fourier Intelligence representative said homes present several distinct challenges. Residential spaces are small and vary widely, placing heavy demands on model generalization and hardware stability. Unexpected situations are common, raising the bar for judgment and engineering robustness. Safety, compliance, and ethics also become more significant constraints when robots operate around consumers.
Multiple industry participants told 36Kr that home robotics will require improvements across the technology stack and that another two to five years may pass before technical capabilities begin to catch up with current public expectations.
Outside the home, however, a number of applications have moved beyond the demonstration stage. In addition to familiar automotive and power-grid inspection scenarios, WRC featured more specialized uses including food packaging, robotic traffic management, and picking for pharmacy and supermarket micro-fulfillment centers, small warehouses designed to fulfill nearby orders.
Galaxea AI demonstrated what it described as the world’s first robot-operated micro-fulfillment center, a system intended for deployment with JD.com. It is designed to autonomously pick items, navigate the facility, and pack orders.
One notable part of the demonstration was bag handling. Deformable objects such as plastic bags have historically been difficult for robots to manipulate precisely. At the booth, the robot opened a bag and continuously adjusted its posture as it packed items. The process still took about two and a half minutes per bag and occasionally required staff assistance.
According to booth staff, the micro-fulfillment setup can process about 100 orders in eight hours and generalize across more than 10,000 SKUs.
Aimoga Robotics, Chery’s robotics company, demonstrated a traffic management application. A staff member said the robots are designed for three main functions: providing legal education and visitor guidance at government service centers and vehicle administration offices; directing traffic signals; and identifying potentially unlawful behavior by riders of bicycles and other nonmotorized vehicles, including carrying passengers illegally, not wearing helmets, traveling against traffic, or running red lights.
The robots can identify license plates and issue guidance, the staff member said, but they do not have law enforcement authority and must work alongside human police officers.
According to the staff member, the robots have already been delivered in Wuhu, Changzhou, Suzhou, and other cities, with about 100 units deployed across ten cities.
Companies from more established robotics segments are also applying their existing capabilities to new tasks. Sunseed continues to focus on food processing, where frequent changes in production requirements make automation difficult but potentially valuable.
A Sunseed staff member said a meat processing line may switch from pork belly to ribs or foreleg meat within an hour, requiring robots to adapt quickly. Food production also imposes strict standards and constant pressure to increase throughput. Those characteristics have kept parts of the sector labor-intensive while creating demand for more flexible automation.
The company said its robots are already producing nearly 10,000 airline meals per day at aviation catering facilities. It has also deployed a meat weighing and packing project for a retailer in Beijing, boxing products such as pork belly while reducing waste.
Keenon Robotics is applying a similar division of labor in hospitality. At one demonstration, a staff member described a laundry room workflow in which a delivery robot transports dirty laundry to a humanoid robot. The humanoid loads and starts the machines, then folds clean clothes into baskets. The delivery robot returns the laundry to guest rooms.
The setup runs on KOM 3.0, a VLA system integrated with a world model, while Keenon ProS is fine-tuned as a domain-specific model for the scenario.
The company describes the setup as a combination of general-purpose and specialized robots. Repetitive, route-based delivery work can be cheaper and more efficient for specialized robots, while flexible operations such as loading machines and folding clothes are assigned to general-purpose humanoid robots. The two categories are designed to complement rather than replace each other.
According to the staff member, the laundry room scenario is undergoing proof-of-concept testing with a large hotel chain.
WRC 2026 offered some evidence that the robotics sector is moving beyond spectacle. Many exhibitors emphasized practical demonstrations over attention-grabbing stunts, while wheeled, articulated, and humanoid robots were increasingly presented as tools for specific jobs.
That does not mean the technology is ready for every scenario. Household robots remain slow and limited, embodied models still need more training data, and some commercial deployments still require human help. But WRC showed that the industry is focusing more on practical engineering, reliable performance, and real-world use.
For manufacturing, those practical questions may matter more than the flashy demonstrations that once drew the biggest crowds.
KrASIA features translated and adapted content that was originally published by 36Kr. This article was written by Qiu Xiaofen for 36Kr.
Note: RMB figures are converted to USD at rates of RMB 6.73 = USD 1 based on estimates as of August 24, 2026, unless otherwise stated. USD conversions are presented for ease of reference and may not fully match prevailing exchange rates.