The Second World Humanoid Robot Games ended after five days of competition, with 1,301 contests completed. The final image was provided by Tiangong Ultra, which crossed the finish line of the 100-meter sprint in 8.64 seconds. That run closed the event and also became a symbol of how far humanoid robots have moved in a short period.
During the five-day schedule, humanoid robots challenged human world records in a series of athletics events, including the 100 meters, 400 meters, and high jump. Beyond the track, 21 scenario events allowed spectators and industry observers to see a more practical vision of the intelligent future. The Games were not only a stage for medals. They were a test of technology, a source of high-quality data, and a map of the difficult points that still need to be solved.
When the curtain fell, humanoid robots had reached a new starting point. The question is no longer whether humanoid robots can perform in controlled demonstrations. The question is how quickly they can become stable, reliable, and useful in the messy, unpredictable real world.
| Area | Recorded detail | Meaning for humanoid robots |
|---|---|---|
| Overall competition | 1,301 contests across five days | A concentrated test of humanoid robots under public pressure |
| 100-meter sprint | 9.39 seconds, 8.85 seconds, and 8.64 seconds from preliminary rounds to the final; Tiangong Ultra finished in 8.64 seconds | Shows rapid iteration in speed, stability, and autonomous control |
| Athletics challenges | Humanoid robots challenged human world records in the 100 meters, 400 meters, high jump, and other events | Pushes humanoid robots toward physical and algorithmic limits |
| Scenario events | 21 events, more than 40 percent of the total | Tests the brain, generalization, and practical intelligence of humanoid robots |
| Tennis demonstration | Fully autonomous human-robot rally between a tennis humanoid robot and Zheng Jie | Validates real-time perception, motion, and human-machine interaction |
| Firefighting scenario | Fire extinguisher weighed four to five kilograms; layout was not disclosed before competition | Exposes physical limits and the need for mechanical adaptation |
| Next technical focus | Visual-language-action models with action chain of thought; world models; hundreds of humanoid robots evolving online simultaneously | Points to long-horizon task stability and scalable learning |
-
Competition Pushes Humanoid Robots Toward Technical Limits
In the 100-meter sprint, the times fell from 9.39 seconds in the preliminary rounds to 8.85 seconds in the semifinals and then to 8.64 seconds in the final. Those numbers mattered because they showed that humanoid robots are not improving in a single laboratory run. They are improving under pressure, in public, across multiple rounds, with competitors and spectators watching. The humanoid robots in the final challenged human world records and stunned the industry.
Tang Jian, chief technology officer of the Beijing Humanoid Robot Innovation Center, said the performance of humanoid robots this year represented a very large improvement over the previous year. He noted that gaps in stability, reliability, and efficiency can be solved, and that the speed of that progress may be faster than many people imagine. The statement was not a prediction based on a single prototype. It was an assessment formed after watching humanoid robots compete at high speed and under repeated stress.
One year earlier, Tiangong was among the few humanoid robots that ran in a fully autonomous mode. This year, it again ran fully autonomously, but at a much higher speed. Tang compared the challenge to riding a motorcycle: when speed becomes very high, control becomes extremely difficult. During turns, the humanoid robot’s speed exceeded 10 meters per second. If control is not precise, the robot can easily be thrown out of balance. For that reason, the team spent significant time solving motion control and navigation problems during turns while also increasing straight-line speed.
The Games have become a real and hard-core training ground for the humanoid robots sector. A competitive environment that is difficult, fully staged, and extreme gives the supply chain a rare stage for large-scale deployment testing and technical iteration. In a normal development cycle, engineers might test one subsystem at a time. In the Games, perception, decision-making, balance, power, thermal control, communication, and mechanical endurance are forced to work together. That integration pressure is exactly what humanoid robots need if they are to leave the laboratory and enter daily life.
The value of the sprint was not limited to a fast time. It showed that humanoid robots can be pushed to the edge of their physical and algorithmic limits in public. It also showed that the edge is moving. A result that once seemed impossible can become a baseline within a year. For the humanoid robots community, the race created a shared reference point and a sense of urgency.
-
Hands and Force Control: Small Components Face Extreme Tests
Hands are one of the most difficult obstacles for humanoid robots entering thousands of industries. A humanoid robot may walk, balance, and navigate, but without a capable hand it cannot perform useful work. The Second World Humanoid Robot Games placed unusual demands on end effectors and force-control systems, turning small components into critical test subjects.
Beijing Lingxin Qiaoshou Company used the diverse event scenarios to test its dexterous hands. Cao Yueyang, the company’s event director, said that beyond the dedicated dexterous hand competition and the pitch-pot event, the champion and runner-up teams in tug-of-war also used its products. The first type of event tested fine manipulation. The second tested durability, stability, and continuous output. Having two extreme performance dimensions validated in one competition was a surprise and a valuable result.
Kunwei Technology provided force-control technology for the Games. Yuan Minglun, co-founder of the company, described the competition as an excellent training stage. It created a test environment with no cost concerns and an orientation toward extreme breakthroughs. Participating teams actively pushed humanoid robots and core components beyond their rated performance limits. Over-rated operating conditions and instantaneous overload appeared frequently. Such scenarios go far beyond ordinary industrial and commercial conditions.
The data collected from those conditions helped the company validate extreme load boundaries, dynamic tolerance boundaries, and fault-tolerance boundaries for high-frequency operations. For humanoid robots, force control is not a secondary feature. It determines whether a robot can hold an object without crushing it, whether it can work alongside people safely, and whether it can recover from contact errors. The Games turned those questions into measurable tests.
The lesson is that humanoid robots improve not only through better artificial intelligence but also through better mechanical honesty. A hand that performs well in a demonstration may fail after hundreds of cycles. A force sensor that works in a clean laboratory may drift under heat, vibration, or unexpected impact. Competition exposes those weaknesses. It forces suppliers to design for the real world, where humanoid robots will encounter friction, wear, delay, and uncertainty.
-
Human Versus Robot: Tennis Match Shows Value of Real-World Confrontation
One of the most memorable moments of the opening ceremony was a world-class human-versus-robot confrontation. A tennis humanoid robot from Galaxy General and tennis star Zheng Jie completed the world’s first fully autonomous human-robot tennis rally. The moment was not a scripted choreography. It required the humanoid robot to perceive a fast-moving ball, predict its path, move into position, and return it in real time.
Wang He, founder of Galaxy General, said that this kind of top-level real human confrontation cannot be replaced by laboratory simulation training. A simulated environment can be controlled. A human opponent cannot. High-intensity real human-robot duels force humanoid robots to face complex and changing dynamic scenes. Through repeated trial and error, the robot exposes technical optimization space, and the team’s future iteration direction becomes clearer.
For humanoid robots, tennis is a demanding test because it combines mobility, perception, timing, and whole-body coordination. The robot must not only see the ball but also understand the human player’s intention, adjust its own position, and produce a controlled stroke. The event showed that humanoid robots can operate in a social and athletic context, not only in a warehouse aisle or on a factory floor.
The tennis rally also illustrated why competition matters for public trust. When humanoid robots perform in front of a live audience, their mistakes are visible. That transparency is useful. It prevents the industry from hiding behind edited videos. It also gives researchers and engineers immediate feedback from a scenario that is close to real human interaction. For humanoid robots to become partners and helpers, they must learn to handle the unexpected movements of people.
The confrontation between a humanoid robot and a professional athlete was therefore more than entertainment. It was a stress test for perception, planning, and control. It was also a reminder that the most valuable training data often comes from environments that cannot be fully scripted. Humanoid robots need those environments if they are to move beyond narrow tasks.
-
Scenario Events Test the Brain and Practical Intelligence
If the athletic competitions tested the physical quality of humanoid robots, the scenario events tested their brains, especially their practical intelligence. The Second World Humanoid Robot Games included 21 scenario events, accounting for more than 40 percent of all events. That was a richer program than the first edition. The shift was from simple single-item competitions to long-process, highly coherent, fully realistic daily-life scenarios.
The tasks included heating food in a microwave, packing medicine for storage, and arranging an office conference table. These are ordinary activities for people, but they are complex for humanoid robots. They require autonomous decision-making, generalization ability, and dynamic adaptability. A humanoid robot cannot simply repeat a fixed motion. It must understand the goal, adapt to the object, and recover from small errors.
Yao Maoqing, co-founder of Zhiyuan Innovation (Shanghai) Technology Co., Ltd., said the core value of the scenario events is to push humanoid robots from a standardized laboratory environment into a real examination room full of uncertainty. In a library scenario, for example, the participating team did not simply stack a large model. Instead, it used high-precision maps and navigation algorithms to let the humanoid robot find its way, then identify labels on book spines and understand their semantics. The real difficulty was not any single technology. It was building the entire system in a stable and reliable way.
The firefighting scenario was even more demanding. The competition layout was not disclosed before the event, and the fire extinguisher weighed four to five kilograms, which created an extreme working condition. The Shanghai Zhiyuan team added a structural component to the humanoid robot’s wrist, changed the action from grasping to hooking, and used the chassis to bear the load. That modification solved a physical bottleneck through mechanical adaptation rather than brute force.
Yao said the team gained something from multiple scenario events. Behind those gains was a reusable embodied intelligence capability base, not a special solution for one track. Different scenarios shared the same base. The greatest value of the scenario events was the continuous polishing of the humanoid robots’ service ability. The adaptability, flexible operation, and autonomous decision-making accumulated in the competition can be transformed into the ability to deploy humanoid robots in industry.
The scenario events also changed how the industry evaluates progress. A humanoid robot that can walk across a stage may impress an audience. A humanoid robot that can find the right medicine, pack it, and place it in the right bin has a clearer path to economic value. The Games made that distinction visible. They showed that the brain of a humanoid robot must be judged by what it can do in a messy room, not only by what it can say in a controlled dialogue.
-
Weaknesses Exposed in Scenario Events Point to Industry-Wide Gaps
The appeal of sports is that it exposes weaknesses. The same is true for humanoid robots competitions. In front of cameras and peers, even teams with proud humanoid robots can make mistakes. That is not a failure of the event. It is a necessary step toward industrial maturity.
In scenario events that were closer to the real world, many humanoid robots performed far worse than humans on simple tasks such as product picking and daily operations. Autonomous humanoid robots did not always perform ideally. This exposed a common shortcoming in embodied intelligence: insufficient generalization ability. Once an object is changed or the environment is slightly altered, the success rate can fall sharply. In other words, humanoid robots still struggle with the accidents that are everywhere in the real world.
Ye Qian, executive committee member and co-president of Shoucheng Holdings, said that this seemingly clumsy performance is precisely the core key to the development of the humanoid robots sector. The public sees the surface functions displayed by humanoid robots. What supports those functions and determines the upper limit of the product is the robot’s ability to adapt. Scenario events focus on the underlying general capabilities of humanoid robots in simulated complex real-world scenes. The ability to achieve applications in multiple real urban scenes shows that technology for humanoid robots has completed a large amount of bottom-layer capability accumulation and has solid underlying support.
Ye believes that the technical iteration from the event will eventually sink down and transform into performance upgrades in civilian products, accelerating the overall progress of the industry. That process will not be linear. It will include failures, retests, and redesigns. But the direction is clear. Humanoid robots must become more flexible, more robust, and more capable of operating without detailed instructions.
The scenario events also revealed that hardware and software cannot be separated. A humanoid robot may fail because its vision system cannot recognize a transparent object, because its arm cannot reach a narrow shelf, or because its planner cannot recover after a small slip. The public may see one mistake. The engineering team sees a chain of dependencies. The Games provided a shared set of failure cases that the entire industry can study.
For the humanoid robots sector, those failures are more valuable than polished demonstrations. They define the next set of problems. They show where investment is needed, where standards are missing, and where human operators may still be required. The industry can only move forward if it accepts that humanoid robots are not yet general-purpose workers. They are developing systems that need many more cycles of real-world learning.
-
Failures Become the Next Roadmap for Humanoid Robots
The problems exposed at the Games have become the main direction for participating teams. Yao Maoqing has already prepared a plan. Improving the brain of humanoid robots will be one of the next technical priorities. The team is trying to let visual-language-action models use an action chain of thought to achieve stable operation in long-horizon complex tasks. It is also using world models to build virtual training environments that conform to physical laws, so that hundreds of humanoid robots can evolve online at the same time and quickly learn new skills in multiple environments and tasks.
This direction reflects a broader shift in research on humanoid robots. Early work focused on making a robot stand and walk. The next stage focuses on making a robot understand a task, decompose it into steps, and adapt when the steps fail. A visual-language-action model can help connect perception to action. An action chain of thought can help the humanoid robot plan over a longer sequence. A world model can provide a safe place to practice before real-world deployment.
At the same time, the competition showed that simulation alone is not enough. Real-world trial and error is a key exploration for the industry as it seeks to break through bottom-layer technical barriers from zero to one. Cao Yueyang said that although humanoid robots still have an obvious gap with humans in practical efficiency in complex scenarios, they have already taken the first step into daily life.
That first step matters. It means humanoid robots are no longer only a laboratory curiosity or a promotional video. They are entering tasks that people recognize: moving objects, following instructions, working in rooms, and interacting with humans. The path is difficult, but it is a path to the future. The teams that left the arena have already prepared to solve these problems.
The humanoid robots sector must also accept that progress will be uneven. Some robots will be better at athletic movement. Some will be better at manipulation. Some will be better at navigation or human interaction. The Games created a rare opportunity for these capabilities to be compared and combined. The failures of one team became a lesson for others. The successes of one component supplier became a reference for the whole supply chain.
-
High-Quality Data Becomes a Strategic Asset for Humanoid Robots
The Games generated more than a spectacle. They generated high-quality data. Humanoid robots operating in extreme athletic events, long scenario tasks, and unpredictable human-machine confrontations produce information that cannot easily be collected in a laboratory. Every fall, every missed grasp, every successful recovery, and every overloaded joint adds to the knowledge base of the industry.
Data is especially important for humanoid robots because the real world is long-tailed. A robot may perform well on common objects but fail on a rare shape, a reflective surface, or a cluttered shelf. The 21 scenario events introduced many such variations. The 1,301 contests over five days created repeated opportunities for humanoid robots to encounter edge cases. The value is not only in the final score. It is in the traces of how the robot reached the score, where it hesitated, and why it failed.
High-quality data also supports better models. Visual-language-action models need demonstrations and corrections. World models need physical interaction data. Force-control systems need contact data. Navigation systems need maps and dynamic obstacle data. The Games provided a structured environment for collecting all of these. Because the events were competitive, teams had strong incentives to document and analyze the data carefully.
For component suppliers, the data helped validate extreme boundaries. For robot developers, the data helped identify weaknesses in perception and planning. For industry observers, the data offered a realistic picture of what humanoid robots can and cannot do today. That picture is more useful than optimistic marketing. It shows where humanoid robots are strong, where they are fragile, and where the next breakthrough may come from.
The strategic importance of data will only grow. As humanoid robots move into homes, hospitals, warehouses, and public spaces, they will need to learn from a vast range of situations. Competitions like the Second World Humanoid Robot Games are not a substitute for deployment. But they are a concentrated form of experience. They compress years of potential learning into a few days of intense activity.
-
From Arena to Ecosystem: Humanoid Robots Move Toward Deployment
The Second World Humanoid Robot Games also revealed an emerging ecosystem around humanoid robots. The event brought together robot developers, component makers, force-control suppliers, dexterous hand companies, algorithm teams, and application partners. A single competition could test a hand in fine manipulation and in tug-of-war. Another could test a force-control system under overload. Another could test a full robot in a library or firefighting scenario. That kind of integration is rare in normal development.
The ecosystem is not only technical. It includes public attention, media coverage, and user expectations. When humanoid robots run, jump, play tennis, and perform household tasks in front of an audience, the public begins to form a more concrete idea of what these machines can do. That awareness can attract talent, investment, and debate. It can also create pressure for safety, reliability, and clear communication.
For the humanoid robots sector, the Games showed that progress depends on multiple dimensions working together. Hardware materials determine strength, weight, and durability. Massive data determines learning. Algorithm models determine intelligence. Computing power determines how quickly models can be trained and deployed. No single breakthrough is enough. The humanoid robots that perform best are those that integrate all these dimensions into a stable system.
The event also showed that humanoid robots are not yet ready for every task. Many scenario tasks remain difficult. Generalization is limited. Autonomous operation is fragile. But the direction is clear. The humanoid robots that competed are not final products. They are research platforms, test beds, and early examples of a new class of machine. Their mistakes are as important as their victories.
As the Second World Humanoid Robot Games closed, the industry entered a new phase. The attention will shift from individual races to long-term capability. The question will be whether humanoid robots can move from a controlled arena to a real environment, from a single demonstration to reliable daily work, and from a collection of impressive parts to a complete, trustworthy system. The teams that competed have already begun that work.
-
A New Starting Point for Humanoid Robots
When Tiangong Ultra crossed the finish line in 8.64 seconds, the Second World Humanoid Robot Games came to an end. But for humanoid robots, the finish line was also a starting line. The five days of competition produced records, failures, data, and new questions. They showed that humanoid robots are improving quickly, but also that the hardest problems remain.
The next stage will require patience. Humanoid robots must become more reliable in unpredictable environments. They must learn to handle objects they have never seen. They must understand human intentions. They must recover from mistakes. They must do all of this while remaining safe and affordable. Those goals cannot be achieved by one event or one company. They require a broad ecosystem and sustained investment.
Yet the direction is becoming clearer. The Games demonstrated that humanoid robots can challenge human world records in athletics. They can play tennis with a professional. They can use dexterous hands in extreme tests. They can attempt realistic tasks in libraries, kitchens, offices, and firefighting scenes. They can generate high-quality data. They can expose their own weaknesses in public. All of these are signs of an industry that is learning quickly.
The road ahead is difficult, but it is a road to the future. The teams that left the arena are prepared to solve the problems they saw. The suppliers are prepared to improve their components. The researchers are prepared to build better models. The public has seen both the promise and the limitations of humanoid robots. That combination of hope and honesty is exactly what a young industry needs.
The Second World Humanoid Robot Games ended with a sprint, but the real race for humanoid robots has just begun. It will be measured not only in seconds and meters, but in reliability, adaptability, and usefulness. Humanoid robots have reached a new starting point. The next chapter will be written in real scenes, with real users, and with the lessons of the Games as a guide.