Fifty Tiangong humanoid robots moved in a steady, orderly formation at the National Speed Skating Oval, maintaining precise spacing and coordinated pace as they circled the venue during the opening ceremony of the second World Humanoid Robot Games. The display appeared simple from a distance: a group of humanoid robots walking together in a controlled procession. Behind the scenes, however, the performance represented a high-intensity test of swarm intelligence, autonomous navigation, real-time correction, and safety management for every humanoid robot in the formation.
Xiangyu Zhang, a navigation algorithm engineer born in the 1990s at the Beijing Humanoid Robot Innovation Center, was responsible for the navigation algorithms used in the entrance formation of the Tiangong humanoid robots. In his view, the formation was not merely a ceremonial march. It was a coordinated exercise in which each humanoid robot had to understand a collective instruction, maintain its own position, follow its own path, keep its own rhythm, and still remain aligned with the larger formation. Each humanoid robot had to listen to the group while also making local adjustments based on what it sensed in its immediate surroundings.
The ability of a humanoid robot to walk steadily depends not only on high-level task planning in the so-called brain but also on motion capability that the industry often compares to the cerebellum. For a humanoid robot, this cerebellum-like capability involves continuous perception of posture and position, real-time correction of gait and speed, and the ability to maintain balance and safe operation when the ground changes, positioning fluctuates, or unexpected interference occurs. Zhang’s work on navigation and coordination is directly aimed at giving humanoid robots a more capable cerebellum.
Last year, Zhang graduated with a master’s degree in transportation engineering from Beihang University and joined the Beijing Humanoid Robot Innovation Center. He saw broad prospects for the humanoid robot industry. Soon after joining, he and his colleagues formed a team and won the 2026 Robot Warrior Challenge. At that time, he was responsible for debugging one humanoid robot. Now he is responsible for the formation debugging of fifty humanoid robots. The shift from one machine to fifty machines captures a wider transition in the humanoid robot field, from single-machine intelligence toward swarm intelligence.
1. A Public Demonstration as a Stress Test for Humanoid Robot Coordination
The opening ceremony formation placed fifty humanoid robots in a shared physical space where every movement had to be visible, stable, and repeatable. The humanoid robots started together, but a single simple command was not enough. Each humanoid robot needed its own position, its own path, and its own rhythm while remaining consistent with the overall formation. The humanoid robots had to understand collective instructions and make local adjustments based on their immediate conditions. This combination of shared direction and local autonomy is central to modern humanoid robot navigation.
Zhang described the formation as a high-intensity group intelligence test. The real challenge was not making fifty humanoid robots succeed once in a demonstration. The real challenge was making them safe, stable, and repeatable in a real venue with fluctuations and interference. A humanoid robot may experience slight positioning variation. Another may respond a little slowly when turning. Another may slow down because of ground conditions, gait status, or safety strategy. A small error in a single humanoid robot can be amplified when fifty humanoid robots coordinate together.
This is why the formation is more than a visual spectacle. It is a practical test of how a humanoid robot perceives itself, how it perceives neighbors, how it estimates its position, how it adjusts speed, and how it preserves safety margins. The humanoid robot cannot simply follow a fixed line and hope for the best. It must continuously confirm its location, evaluate its speed, and judge whether its surroundings are safe. For a humanoid robot, navigation is not a one-time drawing of a route. It is an ongoing process of sensing, correction, and judgment.
The public display also illustrates why humanoid robot engineering is multidisciplinary. The formation involved perception, modeling, control, navigation software, safety strategy, and on-site debugging. Some team members focused on models and perception. Others focused on control and navigation software. The team often described the humanoid robot as a new colleague who needs many people to take care of different capabilities. Some people help the humanoid robot see clearly. Some help it think clearly. Some help it stand steadily and walk well. The result is a humanoid robot system that can turn coordinated code into visible action.
2. The Humanoid Robot Cerebellum and the Work of Navigation
The cerebellum metaphor is useful because it highlights a layer of humanoid robot capability that is easy to overlook. A humanoid robot may have a sophisticated brain for task planning, but it still needs a motion and navigation layer that keeps the body stable and the movement safe. This layer continuously perceives posture and position, corrects gait and speed, and responds to ground unevenness, localization jitter, and sudden interference. Without this cerebellum-like capability, a humanoid robot cannot reliably translate high-level instructions into stable physical movement.
Zhang’s work focuses on navigation and coordination, which are essential parts of this cerebellum for humanoid robots. Navigation tells a humanoid robot where it is, where it should go, and how to move safely. Coordination tells multiple humanoid robots how to share a task, maintain formation, and tolerate local errors. In a single humanoid robot, navigation may involve path planning, localization, speed control, and obstacle avoidance. In a group of humanoid robots, navigation also involves alignment, spacing, timing, and mutual tolerance. The humanoid robot must remain an individual unit and a member of a collective at the same time.
The formation of fifty humanoid robots corresponds to fifty cerebellum-like systems thinking at the same time. If one humanoid robot experiences localization fluctuation, delayed turning, or a deliberate speed reduction due to safety strategy, its cerebellum must complete correction within a short period. It must fine-tune its path, constrain its speed, and re-engage with the formation instead of carrying the error all the way to the endpoint. The quality of the formation depends on whether each humanoid robot’s cerebellum is stable enough to absorb its own error and remain within the shared task framework.
This is a significant shift in humanoid robot design. The goal is not to make every humanoid robot an identical copy with identical behavior. The goal is to make each humanoid robot capable of handling its own local conditions while respecting common rules. A humanoid robot that can sense, correct, and judge continuously is more robust than a humanoid robot that only follows a rigid trajectory. The difference matters in real environments, where ground conditions, localization signals, human movement, and unexpected obstacles can change from moment to moment.
3. From One Humanoid Robot to Fifty: Scaling the Navigation Challenge
The scale change from one humanoid robot to fifty humanoid robots changes the nature of the engineering problem. Debugging one humanoid robot allows engineers to focus on the behavior of a single platform. Debugging fifty humanoid robots requires attention to group behavior, communication, timing, and error propagation. A small delay in one humanoid robot can affect the spacing of neighboring humanoid robots. A small positioning error can create a visible gap or an unnecessary correction. The engineering team must design rules that allow every humanoid robot to remain stable while the whole group remains coherent.
Zhang and his colleagues did not rely on rigid synchronization in which every humanoid robot simply stepped at the same moment. Instead, they created a set of coordinated rules that were unified and flexible. The unified part included task timing, formation targets, and safety boundaries. The flexible part gave each humanoid robot local decision-making authority. Each humanoid robot could make small adjustments based on its actual position, motion status, and immediate conditions. This allowed the humanoid robots to absorb unavoidable fluctuations and interference on a real site. The humanoid robots shared the same general direction while taking care of the small conditions under their own feet.
The formation therefore relied on fifty cerebellum-like systems digesting errors within the same rule set. This is the essence of the transition from single-machine intelligence to swarm intelligence. Single-machine intelligence means one cerebellum continuously perceives, corrects, and judges in a complex environment. Swarm intelligence means many cerebellum-like systems align with one another under the same task and tolerate one another’s local errors. For humanoid robots, swarm intelligence is not just a matter of sending the same command to many machines. It is a matter of designing coordination that remains robust when every humanoid robot has its own perception, its own motion state, and its own safety constraints.
| Humanoid Robot Navigation Stage | Core Requirement | Operational Focus |
|---|---|---|
| Single humanoid robot debugging | Stable perception, localization, and movement for one platform | Path planning, speed control, balance, and safety distance for a humanoid robot |
| Fifty-unit humanoid robot formation | Unified task timing, formation targets, and safety boundaries | Spacing, alignment, local correction, and error absorption across humanoid robots |
| Humanoid robot swarm intelligence | Shared rules with elastic local decision-making | Mutual alignment, mutual tolerance, and repeatable operation under real-site fluctuation |
The table above summarizes the progression described by the engineering work. The humanoid robot does not become reliable simply because it can complete a single motion. It becomes reliable when its navigation and coordination capabilities can scale from one unit to many units without losing safety or stability. For a humanoid robot, scaling is not only a quantitative change. It is a qualitative change in how the system handles uncertainty, local autonomy, and group consistency.
4. Unified Rules and Elastic Local Decisions
The coordination strategy for the fifty humanoid robots can be understood as a balance between unity and elasticity. Unity is necessary because the humanoid robots must share the same task, the same formation target, and the same safety boundary. Without unity, the formation would fragment into isolated movements. Elasticity is necessary because real venues are not perfectly controlled. A humanoid robot may need to adjust its path slightly, reduce speed briefly, or correct its heading to remain safe. If every humanoid robot were forced into a completely rigid trajectory, small disturbances could produce larger errors and less stable behavior.
The design therefore allowed each humanoid robot to make local decisions within a shared framework. The humanoid robot could use its actual position, motion status, and immediate environment to fine-tune its behavior. This is a practical form of humanoid robot autonomy. It does not remove human oversight or central task definition. It gives the humanoid robot enough local authority to digest real-world variation. The humanoid robot remains part of the collective while caring for its own local conditions. The formation becomes more robust because errors are corrected close to their source rather than propagated through the entire group.
This approach also reflects a broader lesson for humanoid robot deployment. In a laboratory, engineers can reduce uncertainty and repeat tests under controlled conditions. In a public venue, a humanoid robot must deal with surfaces, crowds, signals, and unexpected events. The humanoid robot must be able to maintain safety even when conditions are not ideal. Unified rules provide a baseline of safety and coordination. Elastic local decisions provide the flexibility needed to keep the humanoid robot stable when the baseline meets reality.
For humanoid robot swarm intelligence, mutual tolerance is as important as mutual alignment. Each humanoid robot must align with the group’s direction, but each humanoid robot must also tolerate small deviations from other units. The system cannot treat every small deviation as a failure. Instead, the system must allow each humanoid robot to correct its own error while the group continues to move. This is how fifty humanoid robots can maintain a coherent formation without becoming brittle. The humanoid robot formation succeeds when every unit can handle its own local situation and still contribute to the shared task.
5. Young Engineers and the Humanoid Robot Talent Pipeline
The Beijing Humanoid Robot Innovation Center is home to many young engineers born in the 1990s and 2000s. Their daily research is active and energetic, and young people are visible throughout the laboratory. This youthful engineering culture matters for the humanoid robot field because humanoid robot development requires rapid iteration, cross-disciplinary collaboration, and hands-on problem solving. The humanoid robot is not built by a single discipline. It requires mechanical engineering, control engineering, computer science, perception, navigation, artificial intelligence, and systems integration.
Yong Tao, an associate professor at the School of Mechanical Engineering and Automation and the Embodied Intelligence Robot Research Institute at Beihang University, said that the humanoid robot is a multidisciplinary innovation field. Young researchers have contributed to core algorithms, core components, system integration, and scenario exploration for humanoid robots. Universities such as Beihang have opened embodied intelligence programs, but the demand for talent remains urgent. The cultivation and recruitment of the talent team needed for humanoid robot development still require further strengthening. In his view, the young talent team is a core element in the development of embodied intelligence and the humanoid robot field.
Zhang’s team reflects this collaborative model. The members are mostly engineers born in the 1990s and 2000s, and they work together with a high degree of默契, a natural understanding built through close cooperation. The team includes people working on models and perception, as well as people working on control and navigation software. They often describe the humanoid robot as a new colleague who needs many people to care for different abilities. One person helps the humanoid robot see clearly. Another helps it think clearly. Another helps it stand steadily and walk well. This division of labor is essential for a humanoid robot that must operate in complex environments.
The humanoid robot field therefore needs more than individual brilliance. It needs teams that can integrate algorithms, hardware, safety, and real-world testing. It needs engineers who can move between code and physical behavior. It needs researchers who can understand both the high-level task and the low-level motion. The humanoid robot is a system product, and its progress depends on a talent pipeline that can support system-level innovation. The growth of young engineers at the Beijing Humanoid Robot Innovation Center is one sign of how the humanoid robot ecosystem is developing.
6. Code Becoming Real Action in Humanoid Robot Development
Zhang finds a strong sense of contrast in his work. Sometimes he sits in front of a computer looking at maps, trajectories, and data. Sometimes he is on site chasing a humanoid robot to see why it slowed down at a certain place, deviated slightly, or stopped unexpectedly. Many people imagine navigation as drawing a line on a map. In real humanoid robot development, navigation is more like continually teaching the humanoid robot to understand the world. The humanoid robot must learn how to interpret its position, how to move through space, and how to respond to unexpected conditions.
For Zhang, the effectiveness of a piece of code is not decided by the result on a screen alone. It is decided by whether the humanoid robot can complete the task in a real environment. If path planning is improved, the humanoid robot may avoid an unnecessary turn. If speed and safety distance are adjusted well, the humanoid robot may walk more naturally beside people. If exception handling is improved, the humanoid robot may avoid panicking and spinning in place when an unexpected situation occurs. These improvements may appear small in code, but they change the physical behavior of the humanoid robot in meaningful ways.
The feedback loop from code to real action is one of the most attractive parts of humanoid robot research. A change in an algorithm can alter how a humanoid robot moves, how it corrects itself, and how it interacts with people. This feedback loop also makes humanoid robot development demanding. The engineer must understand the code, the hardware, the environment, and the task. The humanoid robot cannot be evaluated only through simulation. It must be tested in the real world, where perception noise, ground variation, human movement, and unexpected events create conditions that no simulation can fully capture.
| Humanoid Robot Development Activity | Practical Purpose | Observable Result |
|---|---|---|
| Path planning optimization | Reduce unnecessary turns and improve movement efficiency | A humanoid robot follows a cleaner route through space |
| Speed and safety distance adjustment | Make movement more natural around people and obstacles | A humanoid robot walks with more predictable and comfortable spacing |
| Exception handling improvement | Prevent unstable reactions to unexpected situations | A humanoid robot avoids panicking or spinning in place |
The table reflects a simple principle in humanoid robot engineering: code quality is ultimately expressed through physical behavior. A humanoid robot that can complete a task in a real environment is more valuable than a humanoid robot that performs well only in a controlled demo. This principle applies to navigation, coordination, manipulation, and safety. For the humanoid robot field, real-world feedback is not an optional step. It is the central test of whether the system works.
7. Continuous Perception and Correction as a Humanoid Robot Capability
In the past, robots more often followed fixed routes and executed predetermined actions. Now, humanoid robot development is paying more attention to continuous perception, continuous correction, and continuous judgment during a task. Zhang often puts himself in the role of the humanoid robot during development. The humanoid robot does not wait until it has gone wrong to turn back. Instead, it continuously confirms its current position, checks whether its speed is appropriate, and judges whether its surroundings are safe. This continuous process is what makes the humanoid robot more adaptive and more reliable.
The improvement of continuous perception and correction technology will accelerate the movement of humanoid robots from laboratories into more application scenarios. A humanoid robot that can perceive and correct in real time is better prepared for industrial environments, commercial service spaces, and home settings. In industry, a humanoid robot may need to move through shared spaces and complete tasks near equipment and people. In commercial service, a humanoid robot may need to navigate around customers and obstacles. In the home, a humanoid robot may need to move through rooms, avoid furniture, and interact safely with family members. Each scenario places different demands on humanoid robot navigation.
Zhang’s current navigation work is aimed at industrial, commercial service, and home scenarios. In the future, a humanoid robot will not merely complete a single action. It will need to understand a task and complete the task as a whole. For example, when given an instruction to deliver an item to a designated location, the humanoid robot must understand where the destination is, determine how to get there, avoid people and obstacles along the way, and complete the handover or operation after arrival. Navigation in this context is like the humanoid robot’s spatial intuition. It tells the humanoid robot where it is, where it should go, and how to move safely.
- The humanoid robot must maintain continuous awareness of its own position and posture.
- The humanoid robot must correct its path and speed in response to local conditions.
- The humanoid robot must align with shared task rules while preserving local safety margins.
- The humanoid robot must translate high-level task understanding into stable physical movement.
- The humanoid robot must operate safely around people, obstacles, and changing environments.
These capabilities are connected. A humanoid robot that perceives well but cannot correct will still drift. A humanoid robot that corrects well but cannot judge safety will still create risk. A humanoid robot that follows a route but cannot understand a task will remain limited. The development of a capable humanoid robot therefore requires an integrated approach. Perception, navigation, control, planning, and safety must work together. The humanoid robot cerebellum is not a single module. It is an integrated capability that supports stable and intelligent action.
8. The Road Ahead for Humanoid Robot Navigation
Beijing has already become part of the first echelon in the development of the humanoid robot and embodied intelligence industry in China. Yong Tao said that in the embodied intelligence and humanoid robot field, Beijing has a complete ecosystem from basic research to innovation and entrepreneurship. This ecosystem is the result of sustained effort. It includes universities, research institutes, companies, testing environments, and application exploration. For the humanoid robot industry, such an ecosystem is important because humanoid robot development requires long-term investment across many disciplines and stages.
In the future, humanoid robots must be able to move autonomously in complex spaces and gradually develop navigation capabilities for more open environments. The goal is to make humanoid robot navigation more intelligent. This goal is not limited to a single demonstration. It points toward humanoid robots that can operate in real settings, understand tasks, adapt to local conditions, and complete work safely. The humanoid robot must be able to move through spaces that are not perfectly structured and respond to conditions that are not fully predictable. This is the direction of humanoid robot navigation research and engineering.
The formation of fifty Tiangong humanoid robots at the National Speed Skating Oval offers a visible example of where the field is heading. The humanoid robots maintained a coordinated formation through autonomous navigation. They demonstrated that a humanoid robot can be part of a larger group while retaining local correction capability. They also showed that humanoid robot swarm intelligence depends on robust individual cerebellum-like systems. The humanoid robot is not only a machine that performs a single action. It is a system that perceives, corrects, judges, and coordinates in real time.
For the engineers involved, the work continues beyond the opening ceremony. Each test, each correction, and each on-site adjustment contributes to a more capable humanoid robot. The humanoid robot must learn to handle real-world variation, maintain safety, and complete tasks in a repeatable way. The path from single-machine intelligence to swarm intelligence is still developing, but the direction is clear. The humanoid robot will need stronger perception, more reliable correction, safer navigation, and better coordination. The humanoid robot that can understand a task and complete it in the real world will be the humanoid robot that creates lasting value.
The development of the humanoid robot cerebellum is therefore not a narrow technical detail. It is a foundation for the next stage of humanoid robot applications. A humanoid robot with advanced navigation can move through industrial sites, commercial spaces, and homes with greater confidence. A humanoid robot with swarm intelligence can work with other humanoid robots and with people in shared environments. A humanoid robot with continuous perception and correction can turn code into real action, step by step, in the physical world. The humanoid robot field is moving forward through exactly this kind of patient, multidisciplinary, real-world engineering.