On a rainy Sunday in Beijing, the second edition of the World Humanoid Robot Games (WHRG) moved its emergency management category out of the showroom and into a working fire brigade. Twenty-three teams registered for the firefighting final, a scenario-based challenge designed to test whether humanoid robots can do more than dance, run, or play music for an audience. According to the event's organizing committee, this year's competition deliberately shifted away from the choreographed demonstrations that defined the first WHRG in 2025, replacing them with what organizers describe as "realistic simulation."
The setting itself was a statement. Instead of a model room with controlled lighting and predictable layouts, the competition took place at an actual fire brigade in Beijing. Human firefighters participated in the exercise, lighting the simulated fire and observing whether each robot performed the extinguisher operation correctly. The task design reflected a specific set of operational requirements: each robot had 30 minutes to complete three distinct tasks. First, the robot had to identify two randomly placed simulated hazardous substances and report their types to judges through returned images. Second, it had to locate three randomly selected open valves of different types and shut them off. Third, it had to identify a fire source, find a fire extinguisher, and spray it until the fire was extinguished.
Of the 12 teams that competed on Sunday, only three finished the challenge. The weather did not cooperate. Rain affected the competition, and shifting outdoor lighting introduced uncertainty into the robots' visual recognition and manipulation systems. These are the kinds of variables that laboratory testing rarely captures, and the results showed the gap between controlled conditions and operational reality.
One team that completed the run within the allotted time was from UniX AI, a company known for developing robot applications for real-world scenarios. UniX AI had recently raised a new round of funding—300 million yuan, approximately $44.49 million, in late March. But even this successful run revealed the technology's limitations. The robot moved noticeably slower than human firefighters would in the same situation. When operating the fire extinguisher, the robotic hand needed two attempts to align with the target before it could begin spraying.
Yang Liqi, a representative of the UniX AI team, explained that the firefighting scenario placed higher demands on both the robot's recognition and manipulation capabilities. Rain and lighting conditions can affect the robot, Yang said, and actual situations are often different from what is simulated in the laboratory. This was one reason a number of teams could not complete their runs on Sunday morning. The UniX AI team had programmed their robot to make up to three attempts when operating the extinguisher. If the robot saw white smoke after the first attempt, it would not make a second attempt—a programmed decision to avoid redundant action when the task was already accomplished.
The difficulty, Yang noted, was not simply whether the robot could move its arm. Humanoid robots rely on different types of joints for different tasks. Small, high-precision joints enable delicate hand operations, while high-torque, reliable joints support movement and balance. A failure at any point in the chain of actions and detection could affect the final result.
Jin Chenran, a representative of the Tiangong team, whose robot only accomplished one of the three tasks, framed the failures as part of the point. Whether the competition goes smoothly or not, that is the significance of a real-world simulation, Jin said. Every real-world failure is valuable data that helps humanoid robots toward practical application. The questions raised by these failures—why the robot failed to recognize an object, how its motion-planning system selected the wrong path—can subsequently be used to improve algorithms and train AI models.
The competition also offered a glimpse into the current transitional stage of humanoid robot technology. At the venue, some team members were wearing VR headsets and remotely adjusting their robots shortly before their runs. Some humanoid robots were seen using omnidirectional wheels instead of feet, while their hands featured multiple joints to enable more precise manipulation. Many teams chose China-developed UBTECH's joint modules for their robots. UBTECH's servo actuators cover a wide torque range from 0.2Nm to 200Nm, allowing robots to combine dexterity with strength. VR teleoperation gives operators a first-person view through the robot's cameras, enabling them to remotely guide movements in real time.
The event sits within a broader push in China to move humanoid robots and embodied intelligence from laboratories and competition arenas into real production and daily-life environments. In June, the Ministry of Industry and Information Technology and other departments launched a special program for humanoid robots and embodied intelligence that encouraged applications in practical fields including emergency rescue.
Zhao Weidong, deputy director of the organizing committee office, said the scenario-based competitions are intended to test the progress of humanoid robots from "competition performance" toward "real operational work," including whether they can eventually become intelligent partners for firefighters. At present, many simulations of actual operations may still be at an early stage, Zhao said. But these are an important starting point for enabling robots to truly assist humans in real-world environments, as well as an important source of data and testing. The ultimate goal, Zhao said, is to push robots into fields where human operations are dangerous and achieve genuine "human-robot complementarity."
Why it matters for European robot service
For European readers tracking the humanoid robot sector, the second WHRG firefighting final is not a distant curiosity. It is a data point about where the technology actually stands, and it carries implications for anyone planning to deploy or service humanoid robots in operational environments.
The most significant takeaway is the gap between demonstration capability and operational reliability. Chinese humanoid robots have long been known for eye-catching abilities such as running, dancing, and music playing. These are impressive feats of coordination and control, but they are performed under predictable conditions. The firefighting competition deliberately removed those conditions. Rain, changing lighting, outdoor environments, randomly placed objects, and randomly selected valves all introduced variables that the robots had not necessarily encountered in the same combination during training.
The results—three out of twelve teams finishing on Sunday—should temper expectations for near-term deployment of humanoid robots in emergency response roles. This is not a criticism of the technology or the teams; it is a realistic assessment of where the field stands. The UniX AI robot that completed the run did so slowly compared to human firefighters, and its manipulation system required two attempts to align the extinguisher. The Tiangong robot completed only one of three tasks. These are not failures in a competitive sense; they are measurements of current capability.
For European buyers and operators, this matters because the humanoid robot market is global, and Chinese manufacturers are major suppliers. UBTECH's joint modules, which many teams chose for their robots at this competition, are already available on the international market. The torque range from 0.2Nm to 200Nm is a specification that European integrators can evaluate for their own applications. But the competition results suggest that the full system—the robot, its perception stack, its manipulation algorithms, and its ability to operate outdoors—is still in a transitional phase.
The use of VR teleoperation is another signal. Some teams were remotely adjusting their robots shortly before runs, using first-person views through the robot's cameras. This indicates that full autonomy in complex, unstructured environments is not yet reliable enough for these teams to trust it without human oversight. Teleoperation is a bridge technology, and its presence at a high-profile competition suggests that the industry recognizes the need for human-in-the-loop control during the transition to greater autonomy.
The policy context is also relevant. China's Ministry of Industry and Information Technology, along with other departments, launched a special program in June for humanoid robots and embodied intelligence, explicitly encouraging applications in emergency rescue. This is a government-level signal that humanoid robots are being positioned for operational roles, not just demonstration roles. European companies and public agencies considering similar deployments should watch how this program evolves, as it may influence the pace of development and the availability of mature systems.
The WHRG's shift from model-room setups to real fire brigades is itself a notable development. It reflects a recognition that laboratory conditions do not adequately represent the complexity of real-world environments. For European robot service providers, this is a reminder that field testing is essential before committing to any deployment. The data generated by real-world failures—why a robot failed to recognize an object, how its motion-planning system selected the wrong path—is precisely the kind of information that improves algorithms and trains AI models. European operators should demand similar testing rigor from their suppliers.
What buyers and operators should know
For organizations considering humanoid robots for emergency response or other outdoor operational roles, the WHRG firefighting final offers several practical lessons.
First, environmental conditions are not secondary considerations; they are primary determinants of performance. Rain affected the competition, and changes in lighting added uncertainty to visual recognition and manipulation tasks. Buyers should ask suppliers how their robots perform in rain, direct sunlight, low light, and other outdoor conditions. If the supplier cannot provide field data from similar environments, that is a risk factor.
Second, manipulation tasks are harder than they appear. The UniX AI robot needed two attempts to align its hand with the fire extinguisher. This is a small but telling detail. The difficulty was not simply whether the robot could move its arm; it was the coordination between perception and manipulation under variable conditions. Buyers should evaluate not just whether a robot can perform a task in a demo, but how many attempts it typically requires in realistic conditions. The UniX AI team programmed their robot to make up to three attempts when operating the extinguisher, with a decision rule to stop if white smoke was seen after the first attempt. This kind of contingency programming is a practical necessity, and buyers should ask about it.
Third, joint architecture matters. Humanoid robots rely on different types of joints for different tasks. Small, high-precision joints enable delicate hand operations, while high-torque, reliable joints support movement and balance. A failure at any point in the chain of actions and detection can affect the final result. The fact that many teams chose UBTECH's joint modules, with a torque range from 0.2Nm to 200Nm, indicates that modular joint systems are becoming a standard building block. Buyers should understand the torque requirements of their specific applications and verify that the robot's joints are appropriately specified.
Fourth, teleoperation is likely to be part of the picture for some time. Some teams at the competition were using VR headsets and remotely adjusting their robots shortly before runs. This suggests that even the most advanced teams do not fully trust autonomous operation in unstructured environments. Buyers should plan for teleoperation capabilities and the associated infrastructure—communication links, operator training, and latency management—rather than assuming full autonomy.
Fifth, the competition results should inform procurement expectations. Of 12 teams that competed on Sunday, three finished the challenge. This is a 25% completion rate under realistic conditions. Buyers should ask suppliers for their own field-test completion rates and compare them to this benchmark. If a supplier cannot provide such data, that is a red flag.
Sixth, the policy environment is shifting. China's Ministry of Industry and Information Technology and other departments launched a special program in June for humanoid robots and embodied intelligence, encouraging applications in emergency rescue. This is likely to accelerate development and may lead to more mature systems in the coming years. European buyers should monitor this program and its outcomes, as it may influence the availability and pricing of humanoid robot systems.
Seventh, the value of real-world failure data should not be underestimated. Jin Chenran of the Tiangong team noted that every real-world failure is valuable data that helps humanoid robots toward practical application. The questions raised by failures—why the robot failed to recognize an object, how its motion-planning system selected the wrong path—can be used to improve algorithms and train AI models. Buyers should ask suppliers how they collect and use field failure data, and whether they are willing to share such data with customers.
Finally, the ultimate goal, as stated by Zhao Weidong, is to push robots into fields where human operations are dangerous and achieve genuine "human-robot complementarity." This is a long-term vision, and the current state of the technology is still early-stage. Zhao acknowledged that many simulations of actual operations may still be at an early stage, but described them as an important starting point for enabling robots to truly assist humans in real-world environments, as well as an important source of data and testing.
Buyers and operators should approach humanoid robot procurement with clear eyes. The technology is advancing, but it is not yet a turnkey solution for emergency response. The WHRG firefighting final provides a realistic picture of current capabilities, and that picture is one of progress tempered by practical limitations. The robots are no longer just dancing; they are attempting real tasks in real environments. But the gap between attempt and reliable completion remains significant, and buyers should plan accordingly.
What is not disclosed in the source material is also worth noting. The article does not specify the names of the three teams that finished the challenge, nor does it provide detailed performance metrics for each robot beyond the general descriptions. It does not disclose the cost of the robots, their battery life, or their maintenance requirements. It does not provide information on spare-part lead times or service-level agreements. Buyers should seek this information directly from suppliers, as it is not available in the public record of this event.
The competition also does not address the economic case for humanoid robots in emergency response. The UniX AI team raised 300 million yuan (approximately $44.49 million) in late March, but the source material does not disclose how that funding translates into unit costs or total cost of ownership. Buyers should conduct their own cost-benefit analysis based on their specific operational requirements.
In summary, the second WHRG firefighting final demonstrated that humanoid robots are moving from demonstration to operational testing, but the transition is far from complete. The technology shows promise, but reliability in real-world conditions remains a challenge. European buyers and operators should use the results of this competition as a baseline for evaluating supplier claims and setting realistic expectations for deployment timelines and performance.
Sources
https://www.globaltimes.cn/page/202608/1368322.shtml
Published by Vigla Media OÜ (Estonia).