BEIJING — The National Speed Skating Oval hosted a scene that looked both choreographed and autonomous. Fifty Tiangong humanoid robots moved with steady steps and orderly ranks, relying on precise autonomous navigation as they circled the venue like athletes during the opening ceremony of the second World Humanoid Robot Games. Behind that display was a team from the Beijing Humanoid Robot Innovation Center, including Zhang Xiangyu, a post-90s navigation algorithm engineer responsible for the entry formation of the Tiangong humanoid robots.
For Zhang, the formation was not merely a parade. It was a demanding test of swarm intelligence. “A formation entrance may look as if the humanoid robots are simply walking forward in an orderly way,” he said. “In reality, it is a high-intensity test of swarm intelligence. When fifty humanoid robots start at the same time, they cannot rely on one simple command. Each humanoid robot must have its own position, its own path and its own rhythm, while still staying consistent with the whole formation. They must understand the collective instruction and also make local adjustments according to what is in front of them.”
The challenge reflects a broader shift in robotics. A humanoid robot must do more than plan a task in its “brain.” It also needs motion ability that industry compares to a cerebellum: the capacity to continuously sense its own posture and position, correct gait and speed in real time, and maintain balance and safety when the ground rises or falls, when localization jitters, or when sudden interference appears. Zhang’s work in navigation and coordination is aimed at giving the humanoid robot a highly developed cerebellum.
Zhang graduated last year with a master’s degree in Transportation Engineering from Beihang University and joined the Beijing Humanoid Robot Innovation Center. He views the humanoid robot industry as having broad prospects. Soon after joining, he and 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 debugging the formation of fifty humanoid robots.
- From One Humanoid Robot to Fifty: The Navigation Challenge Behind a Formation
“The biggest difficulty is that the humanoid robots are not completely identical copies,” Zhang said. “Some may have slight localization fluctuations. Some may respond a little more slowly when turning. Some may actively reduce speed because of the ground, their gait state or safety strategy. A very small error in a single machine is amplified when fifty humanoid robots coordinate. The real challenge is not to make fifty humanoid robots walk successfully once in a demonstration. It is to make them safely, stably and repeatedly complete the task in a real site with fluctuations and interference.”
Fifty humanoid robots moving in formation correspond to fifty cerebellums thinking at the same time. If one humanoid robot experiences localization fluctuation, turning delay, or a voluntary slowdown under a safety strategy, its cerebellum must complete correction within a short time. It must fine-tune the path, gather speed and re-engage with the formation rather than carrying the error all the way to the finish. In other words, whether the formation walks well depends on whether the cerebellum of every humanoid robot is stable enough.
This is not a problem that can be solved by treating the humanoid robot as a simple machine following a fixed line. In a formation, each humanoid robot is both an individual and a member of a group. The individual must protect its own balance and safety. The group must preserve a shared shape and timing. The two requirements can conflict. A humanoid robot that slows down to protect itself may fall out of alignment. A humanoid robot that speeds up to catch the formation may compromise safety. The cerebellum must resolve that tension continuously, not once at the beginning.
The task also changes the meaning of navigation. Navigation for a humanoid robot is not merely drawing a route on a map. It is the continuous process by which the humanoid robot answers three questions: Where am I now? Where should I go next? How can I move there safely? In a single humanoid robot, those questions are already complex. In a formation of fifty humanoid robots, they are asked at the same time, in the same space, under the same task, while the answers of one humanoid robot can affect the conditions faced by another.
That is why Zhang describes the formation as a test of swarm intelligence. A single humanoid robot may perform well in a controlled environment. But when many humanoid robots share a venue, small differences in localization, turning response and speed policy become visible. The formation becomes a stress test for coordination, safety and repeatability. A successful demonstration is not enough. The system must work again and again in the real world.
- What a Formation Entry Reveals About Humanoid Robot Coordination
A formation entry may appear simple to an audience. The humanoid robots move together, keep spacing and follow a route. Yet the visible order is the result of many invisible decisions. Each humanoid robot must know its place in the formation. Each must estimate its own position. Each must compare that position with the formation target. Each must decide whether to adjust speed, path or both. Each must also watch for safety conditions that may require a temporary deviation from the ideal formation.
The humanoid robot therefore cannot be understood as a puppet controlled by one central instruction. It is a moving system with sensors, computation, actuators and safety constraints. The cerebellum of the humanoid robot is the layer that turns perception into correction. It receives information about posture, position and movement. It compares that information with the task and the formation. It then produces adjustments that keep the humanoid robot stable and aligned.
In a small formation, these adjustments may be manageable. In a formation of fifty humanoid robots, they become a collective problem. One humanoid robot’s delay can affect the space available to another. One humanoid robot’s correction can change the local conditions around its neighbors. The formation is not a static shape. It is a dynamic system in which every humanoid robot contributes to the stability of the whole.
That is why the team’s strategy emphasized rules rather than rigid synchronization. The humanoid robots share a common task time, a common formation goal and a common safety boundary. Within those shared limits, each humanoid robot can make local decisions. This allows the humanoid robot to absorb small disturbances without forcing the entire formation to react at once. It also allows the formation to remain coherent even when individual humanoid robots face different local conditions.
- Unified Rules and Flexible Local Decisions
To meet that challenge, Zhang and his colleagues did not make fifty humanoid robots rigidly “sync” every step. Instead, they gave the cerebellum of each humanoid robot a set of rules that is both unified and flexible. The unified part is the task time, the formation goal and the safety boundary. The flexible part is local decision-making. Each humanoid robot can make small adjustments based on its actual position, motion state and immediate situation, absorbing the fluctuations and interference that are unavoidable on a real site.
As a result, the humanoid robots share the same broad direction while also taking care of the small conditions under their own feet. The neatness of the formation comes from fifty cerebellums digesting errors within the same set of rules. This is a subtle but important design principle. If every humanoid robot only follows a central command, the system may be too brittle. If every humanoid robot only follows its own judgment, the formation may lose coherence. The middle path is shared goals with local correction.
Zhang describes this as a process from single-machine intelligence to swarm intelligence. Single-machine intelligence means that one cerebellum continuously perceives, corrects and judges in a complex environment. Swarm intelligence means that many cerebellums align with one another and tolerate one another’s faults under the same task. The humanoid robot is no longer an isolated performer. It becomes part of a network of moving bodies that must remain safe and coordinated.
| Coordination layer | What is shared across the humanoid robot formation | What remains local to each humanoid robot |
|---|---|---|
| Task timing | Common start, common formation goals and common safety boundaries | Adjustment of speed and path based on current position |
| Formation geometry | Overall shape and spacing targets | Local correction when turning response is slower or localization fluctuates |
| Safety | Shared safety limits and stopping rules | Local slowdown or avoidance if ground, gait or safety policy requires it |
The table above summarizes the logic. The humanoid robot formation is not a single brain issuing commands to fifty identical bodies. It is a shared task executed by fifty humanoid robots, each with its own cerebellum and its own immediate view of the world. The shared layer keeps the group moving toward the same goal. The local layer keeps each humanoid robot safe and stable. The combination is what allows the formation to survive real-world variation.
This architecture also explains why the work is difficult to evaluate from a screen alone. A path-planning result may look clean in simulation, but the humanoid robot must still execute it on a floor with changing conditions. A coordination rule may appear mathematically elegant, but the humanoid robot must still apply it while balancing, turning and avoiding obstacles. The cerebellum is where planning meets the physical world. It is where small errors are either corrected or allowed to grow.
- Why the Cerebellum Matters for a Humanoid Robot
For a humanoid robot, stable walking depends not only on task planning in the “brain” but also on motion ability. The cerebellum analogy is useful because it captures the continuous, low-level work that makes movement possible. The humanoid robot must sense its posture and position, correct its gait and speed in real time, and remain balanced and safe when the ground is uneven, when localization shakes, or when an unexpected disturbance occurs. Without that ability, a humanoid robot may understand a task but fail to carry it out.
This is especially important because the humanoid robot operates in a physical environment that does not behave like a clean diagram. The floor may have slight changes. People may move nearby. Sensors may produce small errors. A humanoid robot may need to slow down for safety. Another humanoid robot may need to adjust its path to avoid a collision. The cerebellum must handle these events while keeping the overall task in view. It is not enough to react after a failure. The humanoid robot must continually confirm that its position, speed and surroundings are acceptable.
Zhang often puts himself in the role of the humanoid robot during development. “It does not wait until it has gone wrong to turn back,” he said. “It continuously confirms during movement: Is my current position correct? Is my speed appropriate? Is my surroundings safe?” That continuous confirmation is the heart of the cerebellum. It turns navigation from a one-time instruction into an ongoing dialogue between the humanoid robot and its environment.
As continuous perception and correction improve, they will accelerate the humanoid robot’s movement from the laboratory into more application scenarios. A humanoid robot that can only repeat a fixed motion is limited. A humanoid robot that can perceive, correct and judge during a task can take on more varied work. The cerebellum is therefore not a secondary detail. It is a foundation for practical humanoid robot deployment.
| Stage | Focus for the humanoid robot | Role of navigation and the cerebellum |
|---|---|---|
| Single humanoid robot | One cerebellum perceives, corrects and judges in a complex environment | Keeps balance and completes the route |
| Fifty humanoid robots | Many cerebellums align and tolerate faults under one task | Keeps the formation stable despite local fluctuations |
| Future applications | Understand a task and complete it in open environments | Acts as spatial intuition for safe movement and handover |
The difference between a single humanoid robot and a formation of fifty humanoid robots is not only a difference in number. It is a difference in complexity. In a single humanoid robot, the cerebellum mainly manages the robot’s own body and route. In a formation, the cerebellum must also manage relationships: distance to neighbors, alignment with the group, timing with the task and safety around other moving bodies. The humanoid robot becomes both an independent agent and a member of a collective.
- Why Small Errors Matter in Humanoid Robot Coordination
Small errors are easy to ignore when only one humanoid robot is involved. A slight delay in turning, a small offset in localization or a brief reduction in speed may not change the final outcome. The humanoid robot may still complete its route. But when fifty humanoid robots move together, those small errors can accumulate. They can change spacing, disrupt timing and create local congestion. The formation may begin to look uneven even if no single humanoid robot has failed completely.
The team’s experience highlights this reality. The humanoid robots are not identical copies. Their localization may fluctuate differently. Their turning responses may differ. Their safety policies may cause one humanoid robot to slow down while another continues at normal speed. The cerebellum must recognize these differences and respond locally. It must not wait for a central instruction that may arrive too late. It must correct in the moment, while the humanoid robot is still moving.
That is why the real challenge is repeatability. A demonstration can sometimes succeed because the conditions are favorable. But a useful humanoid robot system must succeed when conditions are not perfect. It must handle small disturbances, recover from minor deviations and continue the task. The cerebellum is central to that ability. It provides the humanoid robot with the continuous correction needed to stay safe and stable.
In the formation, the goal was not to eliminate all variation. That would be unrealistic. The goal was to keep variation within acceptable limits. Each humanoid robot can adjust its path, speed and spacing. Each can absorb local disturbances. The shared rules ensure that these local adjustments do not break the overall formation. The humanoid robot thus becomes part of a resilient system rather than a fragile chain of identical commands.
- The Humanoid Robot Cerebellum and Real-World Safety
Safety is another reason the cerebellum matters. A humanoid robot walking in a public venue must consider not only its own stability but also the safety of people and objects around it. It must maintain safe distances. It must avoid sudden movements that could cause imbalance. It must respond to unexpected obstacles. It must know when to slow down or stop. These decisions cannot be made only at the level of high-level task planning. They must be made continuously during movement.
The cerebellum of the humanoid robot supports this continuous safety process. It monitors posture, position and speed. It checks whether the humanoid robot is following a safe path. It adjusts when the environment changes. If the ground is uneven, it corrects gait. If localization jitters, it filters and corrects. If a disturbance occurs, it works to maintain balance. The humanoid robot remains safe not because nothing goes wrong, but because the cerebellum keeps correcting before small problems become large ones.
In a formation, safety becomes even more complex. The humanoid robots are moving near one another. Their paths must not conflict. Their spacing must remain sufficient. Their speed changes must be coordinated enough to avoid sudden compression or expansion of the formation. The cerebellum of each humanoid robot must therefore be aware of both its own state and the local state of its neighbors. The result is a distributed safety system in which every humanoid robot contributes to the safety of the group.
| Question the humanoid robot cerebellum continuously asks | Why it matters for the humanoid robot |
|---|---|
| Where am I now? | Localization supports formation alignment and safe navigation |
| Where should I go next? | Path planning keeps the humanoid robot aligned with the task and the group |
| How fast should I move? | Speed control balances formation timing with safety |
| Is my surroundings safe? | Continuous perception helps prevent collision and unstable movement |
The table above reflects the ongoing nature of humanoid robot control. The cerebellum does not make one decision and then stop. It keeps asking and answering questions as the humanoid robot moves. This continuous loop is what allows the humanoid robot to handle real-world uncertainty. It is also what allows many humanoid robots to move together without losing formation or safety.
- Young Engineers and the Humanoid Robot Ecosystem
At the Beijing Humanoid Robot Innovation Center, Zhang is not alone. Many young engineers born in the 1990s and 2000s work there. Daily research is full of energy, and young people are visible throughout the laboratories. Tao Yong, an associate professor at the School of Mechanical Engineering and Automation and the Institute of Embodied Intelligence Robotics at Beihang University, says the humanoid robot is a multidisciplinary innovation field. Young researchers have contributed to core algorithms, core components, system integration and exploration of scenario deployment. Beihang University and other universities have opened programs in embodied intelligence, but the demand for talent remains urgent. The cultivation and attraction of the talent team urgently needed for humanoid robot development still needs to be strengthened. In his view, the young talent team is a core element in the development of embodied intelligence and the humanoid robot field.
Zhang says team members are mostly post-90s and post-00s, and they cooperate with strong mutual understanding. The team includes people working on models and perception, as well as people working on control and navigation software. They often say that the humanoid robot is like a “new colleague” who needs many people to take care of it together. Someone is responsible for helping it see clearly. Someone is responsible for helping it think clearly. Someone is responsible for helping it stand steadily and walk well.
Zhang finds a sense of contrast in his work. Sometimes he sits in front of a computer looking at maps, trajectories and data. Sometimes he chases the humanoid robot at the site, checking why it slowed down at a certain place, why it drifted slightly, or why it suddenly stopped. Many people imagine navigation as drawing a line on a map. In real development, it is more like continuously teaching the humanoid robot to understand the world. Whether a piece of code is effective is not finally judged by the result on the screen. It is judged by whether the humanoid robot can complete the task in a real environment.
For example, after optimizing path planning, the humanoid robot may take one less bend. After tuning speed and safety distance, it may walk more naturally beside people. After improving exception handling, it may avoid panicking and spinning in place when a sudden situation occurs. These improvements may seem small individually, but together they make the humanoid robot more capable and more natural. They also show why the cerebellum is so important: it is the layer where code becomes movement.
“This feedback of code becoming real action is what makes humanoid robot research and development particularly attractive,” Zhang said. The statement captures the appeal of the field. A humanoid robot is not only a collection of algorithms. It is a physical system that must act, balance and adapt. When a line of code changes how the humanoid robot moves, the result is immediate and visible. That feedback loop drives young engineers to keep testing, adjusting and improving.
- From Fixed Routes to Continuous Perception and Correction
Zhang notes that in the past, robots more often executed actions according to fixed routes. Now, the focus is on continuous perception, continuous correction and continuous judgment during the task. The humanoid robot is not simply following a script. It is monitoring its own state and the environment as it moves. This shift is central to the development of the humanoid robot as a useful machine rather than a demonstration device.
In this approach, the humanoid robot is always asking questions. Is my current position correct? Is my speed appropriate? Is my surroundings safe? If the answer to any question changes, the humanoid robot must adjust. The adjustment may be small: a slight change in path, a slight reduction in speed, a slight increase in distance from a person or obstacle. But those small adjustments are what keep the humanoid robot safe and effective in a real environment.
Continuous perception and correction also support repeatability. A humanoid robot that succeeds once may have been lucky. A humanoid robot that can succeed repeatedly must have a robust cerebellum. It must handle minor variations without losing stability. It must recover from small disturbances without losing the task. It must coordinate with other humanoid robots without losing formation. That is why Zhang emphasizes that the real challenge is not a single successful demonstration but safe, stable and repeatable completion under real-site fluctuations and interference.
The formation of fifty Tiangong humanoid robots at the National Speed Skating Oval illustrated this principle. The humanoid robots had to maintain a neat formation while navigating autonomously. Each humanoid robot had its own position, path and rhythm. Each also had to remain consistent with the group. The result was not a rigid mechanical march. It was a coordinated movement in which fifty cerebellums worked within one set of rules, each absorbing its own local errors and contributing to the shared goal.
- Navigation as Spatial Intuition for the Humanoid Robot
Beijing has already ranked in the first echelon nationwide in the development of the humanoid robot and embodied intelligence industries, according to Tao Yong. In the field of embodied intelligence and the humanoid robot, Beijing has a complete ecosystem from basic research to innovation and entrepreneurship. That ecosystem is the result of long-term, sustained effort. It provides a foundation for young engineers like Zhang to work on core technologies and move them toward real applications.
Zhang and his colleagues are now aiming their navigation work at industrial, commercial service, family and other diverse scenarios. In the future, the humanoid robot will not only complete a single action. It will truly understand a task and complete the task as a whole. For example, if it receives the instruction “help me deliver an item to a specified location,” the humanoid robot must understand where the destination is, judge how to get there, avoid people and obstacles on the way, and complete a handover or operation after arrival. Navigation in this context is like spatial intuition for the humanoid robot. It lets the humanoid robot know where it is, where it should go and how to walk safely.
This vision expands the role of the cerebellum. The cerebellum is no longer only about keeping balance during walking. It is about supporting the humanoid robot’s movement through complex spaces and enabling it to handle more open environments. The humanoid robot must be able to pass autonomously through complex spaces. It must gradually acquire navigation ability for more open environments. “Making humanoid robot navigation more intelligent is the goal,” Zhang said.
The development path is therefore both technical and human. It requires advances in core algorithms, core components and system integration. It requires scenario deployment and real-world testing. It requires young talent from multiple disciplines. And it requires a clear understanding that the humanoid robot is not a single invention but a system that brings together perception, control, navigation, safety and coordination. The cerebellum is one of the key layers that connects these elements into movement.
- The Humanoid Robot as a New Colleague in the Workplace
The idea of the humanoid robot as a “new colleague” is more than a metaphor. It captures the division of labor inside a robotics team and the future role of the humanoid robot in workplaces. A humanoid robot must see, think, balance and move. No single subsystem can do everything. Perception specialists help the humanoid robot understand its environment. Control specialists help it move smoothly. Navigation specialists help it decide where to go and how to get there safely. Integration specialists bring these abilities together into a working system.
In that sense, the humanoid robot is a collaborative project. It requires many people to care for different parts of its abilities. The cerebellum is one of the meeting points. It takes information from perception, applies control decisions and supports navigation. It is where the humanoid robot’s internal world meets the external world. When the cerebellum works well, the humanoid robot appears calm, stable and natural. When it does not, the humanoid robot may hesitate, drift or stop unexpectedly.
Zhang’s daily work reflects this collaborative reality. He may spend time reviewing data and trajectories. He may also spend time beside the humanoid robot, observing its movement in the real environment. The two kinds of work are connected. Data helps explain what happened. Observation shows why it happened. By moving between the two, the team can improve the humanoid robot’s cerebellum and make its behavior more reliable.
The humanoid robot formation at the National Speed Skating Oval was a public demonstration of this teamwork. It brought together algorithms, control systems, safety policies and young engineers. The humanoid robots moved in formation, but behind them was a larger formation of people working on different parts of the same goal. The result showed how a humanoid robot can move from individual intelligence toward coordinated group behavior.
- The Humanoid Robot and the Future of Embodied Intelligence
The humanoid robot is often discussed as a machine that looks like a person. But the deeper challenge is embodied intelligence. The humanoid robot must understand a physical body, a physical environment and a physical task. It must connect perception to action. It must connect planning to balance. It must connect navigation to safety. The cerebellum is a key part of that connection.
Embodied intelligence means that the humanoid robot cannot be separated from its body and surroundings. A plan that ignores balance will fail. A route that ignores obstacles will fail. A formation that ignores local differences will fail. The humanoid robot must therefore be designed as an integrated system. Its brain, cerebellum, sensors and actuators must work together. Its software must be tested in the real world. Its safety rules must operate continuously, not only at the start.
This is why the work on navigation and coordination is so important. Navigation gives the humanoid robot a sense of where it is and where it should go. Coordination gives it a sense of how to move with others. The cerebellum gives it the ability to correct and adapt. Together, these abilities allow the humanoid robot to move beyond simple demonstrations and toward useful applications.
The formation of fifty Tiangong humanoid robots showed what is possible when these abilities come together. The humanoid robots maintained formation, navigated autonomously and responded to local conditions. They did not move as identical copies. They moved as a coordinated group. Each humanoid robot had its own cerebellum. Each cerebellum followed shared rules. Each cerebellum made local decisions. The result was a formation that could remain orderly in a real venue.
For Zhang and his colleagues, the work continues. The goal is not only to make a humanoid robot walk. The goal is to make a humanoid robot walk intelligently, safely and usefully. The goal is not only to make a humanoid robot complete an action. The goal is to make a humanoid robot understand a task and complete it as a whole. The goal is not only to make one humanoid robot capable. The goal is to make many humanoid robots capable of working together.
That is why the development of a highly developed cerebellum matters. It is the layer that turns perception into correction. It is the layer that turns planning into movement. It is the layer that turns a single humanoid robot into part of a larger system. As navigation becomes more intelligent, the humanoid robot will gain a stronger spatial intuition. It will know where it is, where it should go and how to walk safely. And with each improvement, the humanoid robot will move closer to the goal of understanding a task and completing it in the real world.
The scene at the National Speed Skating Oval was therefore more than a display of fifty humanoid robots. It was a sign of how far the humanoid robot has come and how much further it can go. The steady steps, orderly ranks and precise autonomous navigation were visible to the audience. The continuous perception, correction and judgment were the hidden work of the cerebellum. Together, they showed that a humanoid robot can be more than a machine that moves. It can be a coordinated, adaptive and increasingly intelligent participant in the physical world.