InSpatio Launches Embodied Simulation Data Collection Center in Hangzhou, Plans to Produce 1 Million Hours of Simulation Data per Month

Large language models can directly leverage web pages and code data accumulated on the internet over many years, continuously training to improve model capabilities; autonomous driving companies, meanwhile, can rely on the shadow mode of mass-produced vehicles to continuously collect driving data and human takeover records from real roads. Take Tesla as an example: through millions of mass-produced vehicles worldwide, it integrates data collection into everyday driving, without the need to separately organize large-scale manual collection.

But in the field of embodied intelligence, such data supply conditions do not exist.

Over the past two years, the industry has mainly relied on real-robot teleoperation and first-person (EGO) wearable devices to collect data, and the cumulative effective real-robot data and EGO data obtained amount to only a few million hours. For robotics companies and large-model enterprises that have already invested heavily, obtaining high-quality training data of sufficient scale is still not easy.

English text is being prepared.

With the entire industry spending two years to accumulate a few million hours of data, on what basis does InSpatio plan to raise capacity to a million hours per month? To answer this question, we first need to clearly understand the difficulties currently facing embodied intelligence data collection, and how InSpatio transforms real-world scenes into robot training data.

1. Why can't the current embodied industry achieve large-batch data supply?

A qualified embodied Token needs to simultaneously contain three elements: three-dimensional spatial observation, real physical contact feedback, and executable action sequences.

In other words, the data must not only record what the robot sees, but also include how the robot interacts with objects and what changes occur after the action is taken.

Around these requirements, the industry currently has two main data acquisition paths: one is in the physical world, collecting through real robots and wearable devices; the other is in the digital world, using 3D modeling and physical simulation to generate data. But gradually, both paths have begun to hit efficiency bottlenecks.

1. Hard collection at physical sites: limited by time, environment, and venue

Real-robot teleoperation is currently one of the main ways to obtain robot action data, but collection efficiency is not high.

When an operator controls a robot for one hour of work, after deducting time for equipment debugging, fault handling, and task resetting, the effective collected trajectories may be less than half an hour. To expand the data scale, it is necessary to increase the number of robots and operators, and collection efficiency is greatly limited by physical time and equipment scale.

To reduce investment in robot bodies, many teams have turned to EGO first-person wearable collection, obtaining data by recording human operation processes in real environments, but this method also has limitations.

First, environment adjustment is not flexible enough. EGO collection depends on real sites. If different clutter occlusions, lighting conditions, or countertop materials need to be tested, personnel must be arranged to re-enter the site, move props, adjust the set, and collect again.

Second, many real sites do not allow external teams to modify them at will, such as operating supermarkets, workshops, and pharmacies. Even if problems are found during training, the training team may not be able to return to the original site to reproduce them.

Not to mention high-risk work areas, industrial production lines, dirty and messy environments, security sites, and private residences, which also have strict access restrictions, making it difficult for external teams and equipment to enter.

2. Seeking capacity from the digital world: limited by scene modeling efficiency

Compared with real-environment collection, simulation can repeatedly adjust scenes and tasks in a virtual environment, reducing dependence on real robots and manual labor. But simulation platforms themselves cannot directly provide training data; the first problem to solve is scene modeling efficiency.

Take physical simulation engines such as Isaac Sim as an example. The platform provides a simulated environment for robot operation and physical interaction, but room construction, object modeling, joint binding, collision boundary setting, and physical parameters such as friction coefficients still need manual handling. Finding a 3D artist to build an interactive room from scratch often takes days or even weeks, and both cost and time investment are not low.

So can large models directly complete modeling?

Recently, models such as GPT-6 Astra and Atlas have begun attempting automatic 3D scene generation, but there are still many problems to solve before they can be directly used for robot training.

Take GPT-6 Astra calling Blender for modeling as an example. Its underlying process still requires repeated cycles of "adjust parameters—render into an image—compare with reference image—adjust parameters again." Under this workflow, generating one object takes tens of minutes, and one scene takes several hours. Multiple rounds of inference and rendering further increase generation costs, and procedurally assembled scenes also tend to have an obvious CG feel.

In addition to efficiency and cost, the realism and consistency of generated scenes are also important.

Whether assembling scenes through code parameters or inferring the unphotographed parts of objects from a single image, if real spatial measurements are lacking as constraints, it is difficult to ensure consistency in geometric structure, overall spatial scale, and physical relationships across different viewpoints. These errors will affect the robot's operation in the simulated environment and also increase the difficulty of transferring from simulation to reality, which the industry calls the Real2Sim Gap.

Real-robot and EGO collection are constrained by time, venue, and environment adjustment, while simulation data production faces problems such as low manual modeling efficiency, high AI generation costs, and insufficient scene realism.

How to quickly obtain digital scenes that can support actual robot operation has become a gap that embodied data scaled production urgently needs to cross.

II. InSpatio's Breakthrough Approach: Using Industrial Assembly-Line Logic to Build an Embodied Token Factory

AI computing centers are often called the "Token factories" of the AI era, continuously outputting the tokens required for model training and inference by integrating power, chips, and raw data.

The embodied simulation data collection center that InSpatio has established in Hangzhou adopts a similar approach. Through its self-developed Topos spatial intelligence foundation, it transforms real-world spaces into interactive digital scenes, which are then used to generate robot training data, split into three fully automated industrial production lines:

1. Upstream scene intake: Moving a real site into the cloud in 15 minutes, replacing manual modeling that is measured in weeks

To address the problems described above, InSpatio has changed scene capture to multi-source lightweight input and native 3D generation.

For sites that can be entered briefly, InSpatio uses portable devices for scanning; once the device enters the site, it can complete the capture of hundreds of square meters of space in 15 minutes. For restricted or private locations that are inconvenient to enter, multi-view videos and photos taken by on-site personnel can be used, or native 3D scene generation capabilities can be invoked to construct the corresponding digital environment.

After the raw data enters the Topos engine, the system completes three steps through algorithms:

The first step is object decomposition. The system automatically identifies operable objects such as cups, medicine boxes, drawers, and door handles, and separates them independently from the background, solving the problem in ordinary 3D scanning where objects and scenes easily stick together.

The second step is blind-spot completion. For the backs and bottoms of objects that cameras cannot capture, as well as the hollows left on surfaces after objects are removed, the system invokes native 3D generation models to complete them, generating fully closed 3D models.

The third step is physical calibration. The system generates collision boundaries for objects, assigns physical properties such as mass and friction coefficient, and configures rotation and sliding joint axes for cabinet doors and drawers, enabling them to enter physics simulators directly.

After these processes, real-world scenes such as residences, supermarkets, restaurants, and warehouses can be transformed into runnable digital scenes. According to InSpatio, after scanning is completed, the relevant scenes can be converted within a few hours.

In this way, after digital processing, the same scene can be reused repeatedly in the simulation environment, and object positions, environmental conditions, and interaction methods can also be adjusted to generate different robot training scenes.

2. Midstream compute expansion: Turning on-site prop moving and serial collection into controlled adjustment and parallel synthesis in the cloud

Once a real space is decomposed into an editable digital master, the difficult environment adjustment and single-machine serial bottlenecks that plague EGO and real-robot collection are readily solved.

On InSpatio's production line, once a real scene is stored, it can be called repeatedly. When environmental changes need to be added, the system can directly adjust parameters in the digital scene. After that, the system only needs to introduce a small number of real demonstration actions as seeds to launch multiple simulation channels in the cloud and plan and execute actions in parallel.

The underlying physics engine calculates in real time the mechanical results produced by operations such as grasping, pushing, and placing, and filters out failed attempts such as clipping and slipping, ultimately generating embodied tokens in batches that contain information such as depth, pose, contact force, and action semantics.

This method directly changes the way embodied data is produced: data output will depend more on the reuse level of digital scenes and the scale of cloud parallel computing power. Work that originally required manual repeated setup and collection can basically all be moved to the cloud.

3. Downstream quality control and closed loop: Rejecting homogeneous filler and filling data in a targeted way according to model weaknesses

After data scale expands, how to avoid repeated collection and ensure data effectiveness is also a problem InSpatio needs to solve.

InSpatio has set deduplication standards for scene statistics: all scenes are filed as independent, complete houses or business premises. If only the lighting or shooting angle of the same room is changed, it is counted only as a variant within the scene and is not repeatedly counted in the total number of scenes. On the data delivery side, data is supplied by level according to spatial geometry (L1), physical interaction (L2), and task closed-loop evaluation (L3).

At the same time, data production also forms a closed loop with downstream model evaluation.

When a customer's model exposes weaknesses in testing, for example, frequent failures when encountering transparent cups, highly reflective surfaces, or dense stacked occlusion, InSpatio can, based on the cause of failure, generate in a targeted way a batch of similar complex environments in digital scenes, without needing to send people to the site again to build scenes and collect data. The newly generated data is used to supplement the model's weak links and can also be used for regression verification in the same set of environments.

Conclusion: Treating physical space as raw material, embodied intelligence needs its own Scaling Law

The reason language models can rapidly improve capabilities along the Scaling Law is that the internet completed the digital aggregation of text data on behalf of humanity.

But for robots to leave the laboratory and enter the vastly different real physical world, they always face a data transformation problem—the physical world itself is continuous, disordered, and cannot be directly handed to models for training.

When most teams in the industry are still gradually accumulating data in a linear way by adding camera positions and recruiting more collectors, InSpatio offers another solution: treating physical space as raw material and using 3D reconstruction, generative models, and parallel simulation to build an embodied Token factory that operates around the clock.

When physical interaction data on the scale of millions of hours begins to be scheduled for production by month like industrial goods, large-scale training of embodied intelligence will have a stable data foundation.