The most embarrassing thing for humanoid robots is that they have a pair of legs, but can only stay in a fixed position to work.

They look like humans, but can only dance and perform, and cannot do real work at all.

This is probably the most awkward situation for humanoid robots in the past few years.

Humanoid robots initially attracted massive attention, and the underlying logic is very straightforward: the real world is originally designed around the human body. Stairs, cabinets, shelves, workstations and even door handles all have ready-made sizes. Theoretically, robots can completely use their lower limbs to cross obstacles and directly use their upper limbs to pick up tools to work.

The logic sounds impeccable, but after several years of industry evolution, what we often see is a completely different picture.

The humanoid robots in the videos are mostly running, jumping and somersaulting; even if there are some real work demonstrations, most of them stay on flat ground — walk to a certain spot, stand still, and then reach out carefully.

If the two legs are only responsible for bringing the robot to the task, and it must stand firmly and lock its body before actually operating, then they are just a pair of extremely expensive and extremely fragile mechanical wheels. Where is the advantage of humanoid robots?

It is hardly fair to blame the public for being harsh. Without crossing this threshold, humanoid robots will never be able to shake off the label of "performance prop".

However, solutions to this problem have gradually emerged recently.

Recently, YuanCe Future released a new model and demonstration, taking "Whole-Body Intelligence (WBI)" as the core technical route, collecting human whole-body interaction data by itself, and developing a humanoid foundation model system that can adapt to different bodies.

In the demonstration released by YuanCe, when facing a bicycle lying on the ground, the robot can bend over, exert force with its whole body to lift it up; when facing the start button of the rice cooker, it can complete fine pressing operations.

The part that best reflects the advantage of humanoid robots is a scene of moving a step stool to climb up: when it cannot reach the items at a high place, it moves a step stool by itself, puts it down, then steps onto the platform and grabs the target. This series of actions connects transportation, support switching and high-position operation together.

When the two legs participate in the operation, when the hands cooperate with the balance of the body, the value of the humanoid body begins to appear: robots have the opportunity to use the space designed for humans to complete the work originally undertaken by humans.

Able to walk and jump, why can't they do real work?

In the intuition of many people, since robots can perform difficult actions such as running, jumping and even backflips, it should be easier to take out the trash or wipe the table casually.

The most direct deadlock behind this is that the underlying control system of bipedal robots has always faced a dilemma: to be stable, the hands will shake accordingly; to make the hands accurate, the whole robot is prone to fall.

In the past year, the most mainstream practice in the industry is generally based on "SONIC", the open-source whole-body control benchmark launched by NVIDIA. It has demonstrated whole-body motion tracking, teleoperation, and autonomous operation capabilities connected to VLA. This solution represents the ceiling of academia and industry before, but to further move towards actual operation, it is still necessary to take into account the tracking accuracy of key end effectors, body stability and the stable reachable range at the same time.

One of the key issues is the end effector tracking accuracy.

In YuanCe's comparative test, when the robot stretches towards a low or far place through whole-body coordinated actions such as bending over and stretching forward, the hand end effector tracking error of the open-source baseline is at the order of 10 centimeters. This accuracy is not enough for fine operations.

The deficiency in whole-body control will also propagate forward, limiting the data collection required by subsequent models.

To train a whole-body model capable of operation, the cerebellum is not only responsible for the final deployment and execution, but also acts as the driving base for teleoperation at the front end: when human data collectors perform actions, the cerebellum needs to convert them into robot actions in real time. How much the cerebellum can keep up determines how many of these demonstrations can be turned into effective real-machine data.

One of the obstacles lies in the mapping of large-scale movements.

Picking up the water bottle on the ground and getting the items at a low place require the robot to bend its knees, bend over, and extend the working range of its hands downward. The control system must not only stabilize the body, but also send the hands to the target position. If the posture keeps up but the hands cannot reach, the demonstration still cannot be completed.

Because the robot's body is not isomorphic with the human body, if the data collector touches the ground while the robot's hand is still hanging in the air, and the human has bent to the bottom while the robot is still a short distance away, it performs the action of bending over but does not complete the intention of touching the ground. This not only makes the robot unable to pick up things at the moment, but also makes it difficult for the team to collect effective demonstrations to complete such tasks.

Another obstacle is whether movement and operation can be carried out continuously when the terrain changes.

For mainstream solutions like SONIC, the entire control logic is almost completely bound to flat ground. Its assumption at the algorithm level is that the sole of the foot is in contact with a continuous, smooth two-dimensional plane, which makes it impossible to collect actions under terrain changes.

This also explains the embarrassment in many past humanoid robot demonstrations: they are agile when running and jumping, but when they really need to work, they have to find a flat ground, stand still and then reach out.

Previously, some people in the industry have tried to combine movement and operation, but they adopted a splicing strategy: when going upstairs, run a special movement function, and after reaching the flat ground, switch to the arm operation function.

This is a very awkward transitional solution: if the robot is holding a box and walks to the steps, it even has to drop the box on the flat ground first, then switch to the climbing mode to walk up the steps, which loses the movement advantage of humanoid robots.

If the cerebellum cannot keep up accurately and cannot remain stable under complex postures, the team will find it difficult to collect corresponding real-machine demonstrations; if the model lacks these data, it will be more difficult to learn to continue operating when the body and terrain are constantly changing.

This also explains why humanoid robots can dance better and better, but still stumble when doing work: movements can roughly keep up, but tasks with hands cannot tolerate such deviations.

Use a better S0 model to collect data and train the S1 model

To solve the deadlock of whole-body operation, YuanCe Future has proposed a system architecture called WBI (Whole-Body Intelligence).

The entire system is clearly divided into two layers: the motion control base model "S0" (cerebellum) responsible for high-frequency dynamic balance and motion execution, and the intelligent base model "S1" (brain) responsible for long-range task understanding and decomposition.

The breakthrough first occurred in the underlying S0. The self-developed S0 of YuanCe is responsible for whole-body coordination and stable execution, and implements the task goals into executable actions of the robot.

The most obvious first step is the leap of end effector tracking accuracy.

In YuanCe's test, S0 compresses the hand end effector tracking error to the range of 2 centimeters.

Objectively speaking, 2 centimeters is still not very precise in the broader robot coordinate system, but being compressed within 2 centimeters has reached the pass line of physical contact operation. Obvious progress can also be seen in the demonstration.

After the accuracy, the second difficulty conquered by S0 is the long-standing flat ground defect of SONIC, and it achieves better whole-body tracking at the same time.

S0 expands the control dimension to non-flat ground with height difference. In the process of going up steps, it can also dynamically coordinate the landing trajectory of lower limbs, trunk posture and the smooth transfer of center of gravity.

At the same time, it achieves better mapping for intention tracking of large-scale movements such as bending over.

On the basis that S0 has become a better underlying motion control algorithm, S0 finally feeds S1, the brain with stronger autonomous planning capability.

In the pre-training stage of S1, YuanCe collects EgoBody whole-body human data through the WBI data platform, records the experience of humans moving, adjusting postures, contacting the environment and operating objects in life scenes, and establishes the model's prior knowledge of the relationship between body, movement and environment.

Subsequently, combined with the high-precision demonstrations precipitated by S0 on real machines and DAgger correction experience, S1 has real long-range task understanding capability — in real working conditions, the environment will be continuously reshaped by the robot's own actions, and the result of the previous step must be directly used as the constraint of the next step.

The rice cooker cooking operation of YuanCe Future is a very good case: the robot crosses the height difference of the ground to enter the kitchen, completes rice taking, water adding and inner pot taking and placing at different heights, then closes the lid and presses the button to start, connecting terrain crossing, object transportation and household appliance operation into one task.

It not only shows excellent complex interaction capability, but also shows long-range task capability.

This technical route combining human data, real-machine adaptation and whole-body control has been verified on two configurations of Unitree G1 and Accelerated Evolution T2.

The two robots have different body proportions, joint configurations and motion constraints. In the pantry task demonstration of YuanCe Future, the two robots also cooperate in the refrigerator sorting task.

Unitree G1 and Accelerated Evolution T2 are responsible for refrigerator access and basket holding cooperation respectively, which means that the same technical route has begun to enter the autonomous operation scenarios of different bodies. It still requires body adaptation, but the accumulated capabilities have the opportunity to serve more hardware platforms.

When "Whole-Body Intelligence" becomes the common direction of humanoid robots

For a long time in the past, few companies in the industry could present coherent whole-body task demonstrations.

Recently, domestic and foreign robot companies are intensively demonstrating the new motion capabilities of bipedal robots. From playing ball, to completing some simple tasks, to YuanCe Future's surprising ability to do some housework operations this year, the concentrated emergence of similar demos seems to indicate that humanoid robots are rapidly getting rid of the primary stage of "only being able to dance and perform".

Behind this transformation is the collective maturity of the software and hardware conditions of the entire track: whole-body controllers based on reinforcement learning and simulation are gradually available; first-person head-mounted displays and motion capture devices have greatly lowered the threshold of whole-body data; mature bodies such as Unitree G1 and Accelerated Evolution T2 provide a drop-resistant and durable verification base, and so on.

However, among these emerging achievements, YuanCe Future can go faster and further, largely because they bet all their technical chips on the "Whole-Body Intelligence" route very early.

As a humanoid foundation model company, YuanCe's founding team has a very distinct composite mark of "original academic research + large-scale engineering". In 2026, he became the first Chinese scholar in the history of RSS, the top robotics conference, to win the Early Career Spotlight; Li Tianyu, co-founder and CEO, and Chen Li, Chief AI Officer, are his core partners who have cooperated for many years and are deeply engaged in embodied intelligence and large model architecture.

Li Hongyang promoted "Whole-Body Intelligence" in academia very early. From developing the end-to-end autonomous driving model UniAD to leading the promotion of the million-level real-machine dataset AgiBot World, his core cognition is: modular mechanical splicing will bring fatal information loss and capability fragmentation.

In traditional academia and industry, people working on leg gait do not care about hands, and people working on desktop robotic arms do not care about the lower body. Everyone is discussing dexterous hands or local motion control. But in Li Hongyang's view, these are only local fragments; if hands and feet do not perceive each other, robots will always only be able to "stand at attention" in front of the flat ground workstation. The real general embodied intelligence must move towards "Human Body Learning" — let the whole body learn like a living body from the real whole-body posture and coordination mode of humans, and take the compensation of limbs and center of gravity as the same system for collaborative optimization.

In terms of data worldview, Li Tianyu and Li Hongyang have highly consistent cognition. In a previous in-depth conversation at GeekPark AGI Playground 2026, Li Tianyu once sharply pointed out the blind spot behind the industry craze:

"The time a person actually spends operating in the physical world in his life is only 40,000 hours. Human daily life is highly repetitive and the information value is very low. If the camera cannot see the lower body, it cannot reliably restore the complete whole-body posture; if the key modality is not recorded during collection, it is impossible to retrieve it from millions of hours of videos afterwards."

It is based on this underlying certainty that YuanCe did not take the compromise route of "making a flat-ground robotic arm first and then installing legs under it" like many teams. Instead, from the first day, it regards the whole body as the smallest combat unit: let human data be responsible for the scale and physical common sense of pre-training, and let real-machine data be responsible for the last mile post-training and body alignment, forming a positively operating technical closed loop that connects the S0 controller and the S1 task brain.

Back to the initial question of the article: do humanoid robots have advantages in operation?

In the past, the answers given by many demonstrations were still full of compromises: walk to the flat ground, stop, stand still, and then start to reach out to work.

This set of demos recently released by YuanCe is of course far from the final form of general humanoid robots, but it provides a more specific basis for "footed humanoids can also do work". Two legs can participate in low-position pickup, load transportation and high-position object fetching, and body adjustment can expand the operating range. Those actions that were once mainly used to show flexibility have begun to have the meaning of actual work.

Once this matter is gradually verified, the changes will not only happen in the model. Perhaps in the future, the entire industry will need to redefine this body around "doing work".

The motors at the waist and shoulders must be able to withstand the load of repeated transportation and the heat generated by continuous work; dexterous hands must not only have rich movements, but also withstand high-frequency use; battery life, autonomous charging and battery swapping will also become the key to whether the robot can stay at the work site. The body suitable for running, jumping and performing, and the body suitable for long-time labor may not have the same design priorities.

Software and delivery methods will also change accordingly. The model needs to cover richer experience of contact, terrain and failure recovery, and the body and algorithm teams need to jointly adjust around specific tasks. For customers, what they finally buy should be a set of systems that can continuously complete work on site, and the task success rate, running time and maintenance cost must withstand repeated tests.

This may be the deeper industry significance of "humanoids can also do work": operation capability begins to drive hardware design, data accumulation and system delivery, so that the entire industrial chain is reorganized around actual labor.

When the two legs, trunk and hands jointly become part of completing the task, the body advantage of humanoid robots begins to be realized. And turning this advantage into productivity that customers are willing to pay for is the work that the entire industry needs to complete together next.

This article is from the WeChat Official Account "GeekPark" (ID: geekpark), author: Li Yuan, published with authorization from 36Kr.