Yingsu Launches China's First Embodied Simulation Data Collection Center in Hangzhou, Targeting One Million Hours of Data per Month

The production capacity formula has changed from "people × hours" to "number of environment configurations × parallel computing power."

English text is being prepared.

English text is being prepared.

Training a usable embodied intelligence model requires tens of millions of hours of multimodal interaction data; yet in reality there are only millions of hours, and after removing the low-quality portions, the truly usable amount is even smaller.

A lack of data is a consensus in the embodied intelligence industry.

But fixating only on the word "data" will cause one to miss the more critical issue:

A robot grasping in front of a fixed shelf can have a very high success rate. Move the shelf in a different direction, place an extra row of drinks in front of the target, or reach out from the other side, and the success rate drops. What it has practiced, it can do correctly again and again; what it has not practiced, it does not even know which side to reach its hand from.

What the industry lacks is not just data volume, but also "the variation within the data."

Over the past two years, the mainstream way to fill the gap has been to build venues and add data collection personnel.

According to the China Academy of Information and Communications Technology, as of the end of June this year, more than 70 training venues had been built and put into use nationwide, with another 40 or more under construction or planned.

However, a large number of training venues are concentrated on easy-to-demonstrate tasks such as folding clothes, tidying desktops, and sorting and handling, while the mining of complex long-horizon tasks is severely insufficient. The number of data hours is rising, but the spatial layouts the models have seen are not.

The problem lies in what is being scaled up. Adding venues and adding data collection personnel expands the number of collection runs for the same batch of scenarios—the shelf is still that same shelf, just collected a few more times.

To increase diversity, what is needed is more conditions: layout, occlusion, object combinations, approach paths—each additional one is a new rearrangement.

And how many arrangements a training venue can have depends on how it is set up; change the setup once, and the collection process has to be redone from the start. The labor and time costs of collection multiply several times over.

The more production capacity is added, the more diversity stays in place. What is truly stuck in the industry is not "not collecting enough," but "changing once is too expensive."

Yingsu, from the spatial intelligence track, offers another solution: not only collect a lot, but also exhaust the various permutations and combinations of space, collecting "differences."

On October 29, Yingsu officially launched the country's first simulation-based embodied intelligence data collection center in Hangzhou, with the goal of raising production capacity to the level of one million hours of data per month.

With 3D scene reconstruction and generation at its core, it connects real space, generative models, and simulation capabilities together to build an editable, reusable, and scalable environment and data supply system.

To raise the banner of "simulation" amid the consensus that "real data is king" requires confidence.

This confidence comes from Yingsu directly transforming the data collection method while simultaneously resolving the two major "bottlenecks" of data diversity and production capacity; it also comes from the fact that, as a spatial intelligence foundation model company itself, it knows best what data models need and what truly counts as high-quality data.

To understand Yingsu's choice, one must first see clearly where the costs of the data collection process it aims to replace are actually spent.

EGO first-person perspective collection and teleoperation are the two main ways for embodied intelligence to obtain real behavioral data. EGO requires personnel to enter the environment and complete operations; teleoperation requires operators to participate continuously and adapt to specific devices and tasks. The two methods differ, but neither can avoid "people."

If you cannot avoid people, you cannot avoid organization. Adding a new spatial layout or supplementing a type of operating condition means the process has to be run through again: find a venue, book a time slot, gather the people, set up the equipment, and dismantle it after collection.

In addition, the task proficiency, operating habits, and degree of adaptation to different devices of collection personnel also directly affect collection efficiency and data consistency.

Reorganizing the process is not only cumbersome, but also cannot be parallelized. Venues must be negotiated one by one, work can only start once people are gathered, and once a scenario is collected and dismantled, it is gone.

Thus, every additional variation requires paying the full organizational cost once again.

This is the real reason diversity cannot rise: it is not that the industry does not want to add conditions, but that every addition requires reorganizing from scratch, with costs stacking linearly.

What Yingsu aims to do is bring this cost down while pulling data diversity up.

In a simulation environment, spatial layouts, object positions, occlusion relationships, and task conditions can be expanded through generation, editing, and combination, without having to rearrange physical venues and personnel operations for every variation. The same task can be placed into different environmental conditions, covering variations that are difficult to exhaust in real collection.

For algorithm teams, it is a space for expanding the training distribution and reproducing failure conditions; for robot hardware manufacturers, it is an environmental foundation that can be repeatedly configured and validated.

The simultaneous improvement of data diversity and production capacity also comes from parallel computing across the entire data production process.

Based on a small number of real-machine or teleoperation demonstration seeds, the system combines motion planning, policy driving, and automated physical quality inspection to automatically synthesize, in batches across hundreds of parallel environments, closed-loop trajectories aligned with RGB, depth, segmentation, 6D poses, and action instructions. Environments can be reused, conditions can be adjusted in batches, and the scale of data production thus expands with computing power and the task system, rather than growing only linearly with the number of people.

The divide between the two paradigms lies here: over the past two years, the industry has competed over who has more venues and more collectors, with the production capacity formula being "people × hours"—hours rise, but the condition space does not; the formula of the Yingsu data collection center is "number of environment configurations × parallel computing power"—each additional configuration adds a set of condition coverage, and diversity expands along with computing power.

Based on this system, the Yingsu data collection center is expected to achieve a data production capacity of one million hours per month (or one hundred thousand hours per day)—which means that the effective output of a single center's short-term operation alone can surpass the combined total of the millions of hours of existing stock accumulated by the entire industry over the past two years.

What matters more than these numbers is the innovation in how work is done that they point to: whichever type of occlusion, layout, or operating condition the model fails under, environments are expanded, samples supplemented, and regression tests run specifically for those conditions, shortening the cycle from discovering a problem to verifying it again.

The tighter data production is coupled with R&D, the more direction there is for "how to produce high-quality, high-demand data," and the fewer detours are taken down the path of "collecting a huge amount but being unable to use it."

02

Real spaces set the anchor, simulation generation scales up volume

But in the embodied intelligence industry, simulation has always faced an unavoidable doubt: the environments are not realistic enough.

This is the old Sim-to-Real problem, and it is also where the confidence behind the saying "real data is king" comes from.

Yingsu did not sidestep it; instead, it broke the problem apart: geometry, appearance, and spatial relationships are anchored by real collection, while expansion and repeated verification are left to simulation.

Thus scene assets have two input pipelines.

One comes from real spaces.

Through deep collaboration with Hesai Kosmo, Yingsu connects the hardware and reconstruction capabilities on the collection side with its own model-side 3D dimensional upgrading and physics simulation capabilities.

The strength of Kosmo, this 3D world data engine, lies in simultaneously lowering both the time cost and the precision threshold of turning "real scenes into computable 3D assets": an ordinary 100-square-meter indoor space can be scanned in about 15 minutes, 3DGS reconstruction reaches millimeter-level precision, and one pass outputs three types of aligned data: 3DGS, mesh, and depth.

English text is being prepared.

In other words, the real-world data collected by Kosmo provides a reliable reference for scene geometry, appearance, and spatial relationships; Yingsu then turns these scene data into editable, reusable, and extensible simulation environment assets. Real collection, 3D reconstruction, and world model training—this chain is thus fully connected.

Another pipeline relies on generative models to expand supply.

Yingsu's launched world simulator Topos Pro has the core capability of turning any spatial input into an object-level world that is understandable, editable, interactive, and simulatable.

It supports multi-source inputs such as single images, multiple images, video, LiDAR, and drone aerial photography, automatically completing segmentation of scene objects, alignment with spatial geometry, and assignment of physical attributes. More importantly, the tables, chairs, walls, and equipment within it can be individually selected, replaced, and adjusted, and assigned attributes such as collision, movable, and contactable.

No matter how much video there is, models can only watch; a world that can be acted upon, touched, and verified is what embodied training truly lacks. What Topos Pro does is turn "watching" into "interactive" and "editable."

The two pipelines have a clear division of labor: real-scene reconstruction is responsible for pinning down "whether it looks right," while generative capability is responsible for filling in "whether there is enough." Materials from a single collection are no longer used only once; images and videos on the internet can also be dimensionally upgraded into usable 3D scenes.

Unlike other data collection methods that "collect once, use once, change the layout once, and run through the process again," Yingsu uses real collection to set anchor points, and Topos Pro reorganizes the same collection into countless layouts, occlusions, and object combinations; each reorganization is a new coverage of conditions.

Scene assets thus change from "consumables" into "reusable raw materials": collect once, and it is not only used once, but grows into a space of conditions that can be recycled indefinitely.

Once the scene data is in place, there is still a key processing step: making the things in the scene individually operable. This step is the watershed determining whether a scene can change from "a consumable that is torn down after one collection" into "an asset that can be repeatedly configured."

Take "taking a specified product from a shelf and then placing it into a turnover box" as an example. The system needs to preserve the spatial relationships among the shelf, the product, and the container, isolate the target product, complete the occluded parts, and repair the background gaps left after extraction.

Specifically, the system performs three steps of processing:

Step one, instance-level 3D decoupling. The target product is peeled out of the shelf grid to become an object that can be independently selected, moved, and configured, while preserving its spatial constraints with the shelf and adjacent items. Precise decoupling is the prerequisite for subsequent object-level re-editing.

Step two, occlusion completion and 3D background repair. Complete the occluded back geometry and repair in real time the background holes left after the product is removed, so that when objects are moved or replaced, there will be no geometric gaps. This step determines whether a scene can "withstand" repeated editing.

Step three, task-level physical attribute calibration. Calibrate mass, center of mass, and friction coefficient according to task requirements, and finely process the geometry of the object's bottom, contact surfaces, and nearby obstacles, so that contact judgments in simulation hold up.

After these three steps of processing, researchers can change product positions, add occlusions, adjust placement areas, and observe the robot's performance under different conditions. One scene input can thus serve multiple rounds of environment editing and task experiments.

Yingsu's technical advantages are precisely embodied in these specific capabilities: transforming multi-source inputs into 3D scenes, enabling key objects in the scene to be edited and interacted with, and then configuring and verifying them in combination with target robots and tasks, so that spatial content truly enters the R&D process of embodied intelligence.

This step responds to the industry's doubts about "simulation data" and Sim-to-Real. Whether simulation is realistic enough is solved through concrete structure: geometry, appearance, and spatial relationships come from real capture, physical properties are configured and calibrated, and simulation builds on top of this for expansion and repeated verification; but simulation is not the endpoint—policies ultimately still need to be tested and validated on real machines.

In other words, simulation does not directly "replace" real capture; rather, the structured scene understanding established by real capture is "amplified" by simulation: amplified to the point where it can cover long-tail scenarios, be repeatedly configured and verified, and enter the R&D process of embodied intelligence at scale.

For assets to be usable, they also need to be selectable. For algorithm vendors and robot body manufacturers, the value of a data collection center ultimately comes down to usable environments and data.

Different tasks have different requirements for environments. Yingsu has defined an industrialized "three-tier environment delivery standard" for different task types:

L1 Mobile navigation level, delivering large-scale spatial topology, traversable areas, and dynamic interference, corresponding to "can it walk through."

L2 Dexterous grasping level, delivering millimeter-level object geometry, collision bodies, and dense occlusion combinations, corresponding to "can the hand reach in and grasp it."

L3 Contact and articulated dynamics level, targeting complex interactions such as opening drawers and plugging/unplugging, delivering joint damping, contact forces, and dynamics calibration, corresponding to "how much force is used, and will it get stuck."

The three tiers represent increasing precision and physical fidelity. This distinction has practical significance: in the industry, "high-quality data" has long been an adjective—everyone says what they produce is high quality, but there is no way to verify it.

After tiering into three levels, which tier the geometric precision of the data reaches, how collision bodies are defined, and what range the damping calibration falls within can all be clearly reflected on the delivery sheet, making it easier for downstream teams to select resources suited to their own tasks.

There is another point worth noting: the higher the tier, the harder it is for real capture scenarios to cover, and the greater simulation's relative advantage becomes. At L1, on-site scanning can still scale up volume; at L2, the permutations and combinations of dense occlusions can only be assembled in physical spaces by rearranging again and again; at L3, each drawer's damping and each contact's mechanical response are different, making exhaustive enumeration nearly impossible.

Therefore, task tiering delineates not only standards, but also the range in which simulation can truly deliver value.

Reuse of scenes and assets can reduce the work of repeatedly building environments; continuously supplementing layouts, objects, and task conditions can support model testing, failure reproduction, and policy comparison. At the same time, customer usage feedback can become the basis for the next round of data production.

At this point, Yingsu has achieved a closed loop across the entire data collection process: real capture provides the reference, scene reconstruction turns it into assets, generation and simulation expand coverage, task verification provides feedback, and R&D feedback in turn guides the next round of production. The real world, generative models, and simulation capabilities each play their role within this loop.

The significance of this loop lies not only in the process itself, but also in further accelerating production speed: once data production can absorb R&D feedback, iteration no longer means waiting round after round for collection schedules, but instead directly consolidates into a continuous action of "discovering problems, expanding conditions, and re-verifying."

05

Building an open ecosystem: where do scene resources come from

At this point, building a simulation data collection center still faces a more practical problem: no matter how many environments there are, there must be sufficient scene resources.

Therefore, coordination among scene resources, collection organizations, technical capabilities, and application needs is crucial. This is not a chain that one company can complete behind closed doors.

Based on this, Yingsu chose to cooperate with Hangzhou Meihui Interactive Data Collection Technology Co., Ltd. As a specialized embodied data collection force incubated within the ecosystem of Daily Interactive, a leading A-share data intelligence company, Meihui Interactive inherits Daily Interactive's deep accumulation—a leading domestic data intelligence listed company in data inclusion rate—in large-scale device coverage, DiOS data governance, and a public service network covering more than 2,900 districts and counties nationwide.

In other words, what Yingsu obtains is not ready-made data production capacity, but an entry point that can bring downstream needs forward into the collection design stage.

The value of this entry point lies not at the delivery end, but at the design end. Meihui Interactive's real-scene collection and organization capabilities are connected with Yingsu's 3D reconstruction, generation, and simulation capabilities, driving the transformation of real data into reusable environments and task assets.

What tasks downstream needs to verify, which spatial relationships to preserve, and which environmental changes to supplement can all enter the collection and generation process earlier, rather than discovering problems only after data delivery. The "editability" of scenes is no longer an "after-the-fact remedy," but directly becomes an "input condition" on the first day of the production process.

But scenes brought in by a single partner cannot support the breadth a model needs. Yingsu had been thinking about this earlier.

As early as August this year, it joined hands with more than 20 industry-academia-research units including Hesai, Unitree, Deep Robotics, SenseTime, and Huace Film & TV to jointly launch the "Large-Scale Spatial Intelligence 3D Data Open Plan (SIDO)," carrying out co-construction around spatial data and evaluation benchmarks.

What SIDO aims to solve is not "collecting more data," but dismantling and rearranging the structure of data supply. No single source can simultaneously satisfy breadth, authenticity, and long-tail coverage, so what SIDO explores is three complementary data paths:

First, real-world collection. Industry partners collect real scenes on site through LiDAR, multi-view imagery, robots, and other sensors, managing the anchor points.

Second, upgrading internet content into higher dimensions. Through reconstruction and generation technology, publicly available images and videos on the internet are upgraded into 3D/4D scenes, managing breadth.

Third, simulation and synthesis. Through simulation, extreme and long-tail scenarios that are difficult to collect in reality are synthesized, managing the generalization boundary.

The three paths are not parallel alternatives, but interlock with one another: without real-collection anchoring, upgraded data has no physical reference; without upgrading for volume expansion, real collection cannot support diversity; without simulation to fill in the long tail, the first two combined would still miss the very conditions most in need of generalization.

No single company can complete this structure alone. Yingsu recognized this early on, so it took the lead in industry co-construction and cooperation, while at the same time solidifying the simulation line it excels at most, using reusable simulation capabilities to fill in the long-tail conditions that neither real collection nor upgrading can reach, and then contributing this capability into the co-construction framework.

Not only that. To keep scale always aligned with the model's real gaps, the accompanying evaluation Benchmark system defines standards in reverse: it is not about collecting as much as possible, but about using evaluation results to tell the next round what to collect. Whatever conditions a model fails under, the next round supplements that type of data; the more data there is, the more it aligns with failure points, rather than becoming more scattered the more is collected. Leiphone Leiphone Leiphone

When the productivity formula for embodied data collection switches from "people × hours" to "number of environment configurations × parallel computing power," a closed loop has taken shape in which real collection sets the anchor points, the Topos Pro world model expands scenes, and high-fidelity simulation produces scaled-up interactive experience output.

What Yingsu has landed in Hangzhou is not only a simulation data collection center, but also a computing engine that drives embodied intelligence away from the "manual workshop" and toward general-purpose generalization.