Author | Wang Han Editor | Moying
Over the past few years, the two waves of large-scale AI development have both relied on abundant data supply. Large language models can directly utilize web pages and code accumulated on the internet over many years, improving model capabilities through continuous training. Autonomous driving companies can rely on the Shadow Mode of mass-produced vehicles to continuously collect driving data and manual takeover records from real roads.
Taking Tesla as an example, it integrates data collection into daily driving processes through millions of mass-produced vehicles globally, eliminating the need for additional large-scale manual collection efforts. However, in the field of embodied intelligence, such data supply conditions do not exist.
In the past two years, the industry has primarily relied on real-machine teleoperation and first-person (EGO) wearable devices for data collection, accumulating only several million hours of valid real-machine and EGO data. For robotics companies and large model enterprises that have already invested significant capital, obtaining high-quality training data at a sufficient scale remains challenging.
On October 30, spatial intelligence company InSpatio launched China's first embodied simulation data collection center in Hangzhou. InSpatio expects this center to produce 1 million hours of simulation training data per month, with the comprehensive cost per data record being approximately 1/10 of traditional real-machine teleoperation collection.
The entire industry spent two years accumulating several million hours of data. On what basis does InSpatio plan to boost production capacity to one million hours per month? To answer this question, we must first understand the difficulties currently facing embodied intelligence data collection and how InSpatio transforms real-world scenarios into robot training data.
01.Why Can't the Current Embodied Industry Achieve Mass Data Supply?
Qualified embodied Tokens need to simultaneously include three elements: 3D spatial observation, realistic physical contact feedback, and executable action sequences. This means 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 performed.
Around these requirements, there are currently two main data acquisition paths in the industry: one involves collecting data via real machines and wearable devices in the physical world; the other generates data using 3D modeling and physics simulation in the digital world. However, both paths are gradually encountering efficiency bottlenecks.
1. Hard Collection in Physical Sites: Limited by Time, Environment, and Venue
Real-machine teleoperation is currently one of the primary methods for obtaining robot action data, but collection efficiency is low. When one operator controls one robot for one hour, after deducting time for equipment debugging, fault handling, and task resetting, the effective collected trajectory may be less than half an hour. Scaling up data requires increasing the number of robots and operators, meaning collection efficiency is heavily constrained by physical time and device scale. To reduce investment in robot hardware, many teams have shifted to EGO first-person wearable collection, recording human operations in real environments to obtain data. However, this method also has limitations.
First, environmental adjustments lack flexibility. EGO collection relies on real venues. If testing different clutter occlusions, lighting conditions, or table surface materials is required, personnel must return to the site, move props, adjust setups, and recollect data.
Second, many real venues do not allow external teams to make arbitrary modifications, such as operating supermarkets, workshops, and pharmacies. Even if issues are identified during training, the training team may not be able to return to the original venue for reproduction. Furthermore, scenarios like high-risk work zones, industrial lines, dirty or chaotic environments, security-sensitive areas, and private residences have strict access restrictions, making it difficult for external teams and equipment to enter.
2. Seeking Capacity in the Digital World: Limited by Scene Modeling Efficiency
Compared to real-environment collection, simulation allows repeated adjustment of scenes and tasks in virtual environments, reducing reliance on real machines and manpower. However, simulation platforms themselves cannot directly provide training data; the first issue to solve is scene modeling efficiency.
Using physics simulation engines like Isaac Sim as examples, these platforms provide simulated environments for robot operation and physical interaction, but room construction, object modeling, joint binding, collision boundary setting, and physical parameters like friction coefficients still require manual processing.
Hiring 3D artists to build an interactive room from scratch often takes days or even weeks, resulting in high costs and time investments.
Can large models complete the modeling directly?
Recently, models like GPT-6 Astra and Atlas have begun attempting automatic 3D scene generation, but there are still many issues to resolve before they can be directly used for robot training.
For instance, when GPT-6 Astra calls Blender for modeling, its underlying process still requires iterating through "adjust parameters—render image—compare with reference image—readjust parameters." Under this workflow, generating one object takes tens of minutes, while one scene takes hours.
Multi-round inference and rendering further increase generation costs, and procedurally stitched scenes often exhibit obvious CGI artifacts. Besides efficiency and cost, the realism and consistency of generated scenes are crucial. Whether stitching scenes via code parameters or inferring unphotographed parts of objects from single images, lacking real-space measurement constraints makes it difficult to ensure geometric structure, overall spatial scale, and physical relationships remain consistent across different viewpoints.
These errors affect robot operations in simulation environments and increase the difficulty of sim-to-real transfer, known in the industry as the Real2Sim Gap.
Real-machine and EGO collection are constrained by time, venue, and environmental adjustments, while simulation data production faces issues such as low manual modeling efficiency, high AI generation costs, and insufficient scene realism. How to quickly obtain digital scenes capable of supporting actual robot operations has become a critical hurdle for scaling embodied data production.
02.InSpatio's Breakthrough Approach: Building an Embodied Token Factory with Industrial Assembly Line Logic
AI computing centers are often called "Token Factories" of the AI era, integrating electricity, chips, and raw data to continuously output Tokens required for model training and inference. The embodied simulation data collection center launched by InSpatio in Hangzhou adopts a similar approach. Through its proprietary Topos Spatial Intelligence Foundation, it transforms real spaces into interactive digital scenes for generating robot training data, dividing the process into three fully automated industrial production lines:
1. Upstream Scene Intake: Transferring Real Venues to the Cloud in 15 Minutes, Replacing Weekly Manual Modeling
To address the aforementioned problems, InSpatio changed scene collection to multi-source lightweight input and native 3D generation. For venues allowing brief entry, InSpatio uses portable devices for scanning, completing the collection of hundreds of square meters within 15 minutes of device deployment. 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 corresponding digital environments.
Once raw data enters the Topos engine, the system completes three steps via algorithms:
The first step is object decomposition. The system automatically identifies operable objects such as cups, medicine boxes, drawers, and door handles, separating them independently from the background to solve the issue of object-scene adhesion common in ordinary 3D scans.
The second step is blind spot completion. For object backs and bottoms not captured by cameras, as well as voids left on surfaces after moving objects, the system invokes native 3D generation models to fill in gaps, creating complete closed 3D models.
The third step is physical calibration. The system generates collision boundaries for objects, assigns physical attributes like mass and friction coefficients, and configures rotational and sliding joint axes for cabinet doors and drawers, enabling them to directly enter physics simulators.
After these processes, real-world scenarios such as residences, supermarkets, restaurants, and warehouses can be transformed into runnable digital scenes. According to InSpatio, after scanning, related scenes can be converted within a few hours. Thus, after digitization, the same scene can be reused repeatedly in simulation environments, and object positions, environmental conditions, and interaction methods can be adjusted to generate different robot training scenarios.
2. Midstream Computing Expansion: Transforming On-Site Prop Arrangement and Serial Collection into Controlled Cloud Adjustments and Parallel Synthesis
Once real spaces are decomposed into editable digital masters, the most troublesome issues for EGO and real-machine collection—difficulty in environmental adjustment and single-device serial bottlenecks—are resolved. On InSpatio's production line, once a real scene is warehoused, it can be repeatedly invoked. When environmental variations are needed, the system can directly adjust parameters in the digital scene. Subsequently, the system only needs to introduce a small number of real demonstration actions as seeds to launch multiple simulation channels in the cloud, planning and executing actions in parallel.
The underlying physics engine calculates mechanical results from operations like grasping, pushing, and placing in real-time, filtering out failed attempts such as mesh penetration or slipping, ultimately batch-generating embodied Tokens containing information such as depth, pose, contact force, and action semantics. This method directly changes the production mode of embodied data: data yield will depend more on the reusability of digital scenes and the scale of cloud parallel computing power. Work originally requiring repeated manual setup and collection can basically be transferred to the cloud for completion.
3. Downstream Quality Control and Closed Loop: Rejecting Homogenized Padding, Targetedly Supplementing Data Based on Model Weaknesses
As data scale expands, avoiding duplicate collection and ensuring data validity are also problems InSpatio needs to solve. InSpatio has set deduplication standards for scene statistics: all scenes are archived as independent, complete houses or business premises. Changing only lighting or shooting angles within the same room counts as internal variants of the scene and is not counted towards the total number of scenes.
On the data delivery end, supply is graded according to spatial geometry (L1), physical interaction (L2), and task closed-loop evaluation (L3). Meanwhile, data production forms a closed loop with downstream model evaluation.
When customer models expose weaknesses during testing—for example, frequent failures when encountering transparent cups, highly reflective surfaces, or dense stacked occlusions—InSpatio can targetedly generate a batch of similar complex environments in digital scenes based on failure causes, without needing to send personnel back to the site to set up scenes and collect data. Newly generated data can supplement model weak points and undergo regression validation in the same environment.
03.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 for humanity. However, as robots move out of laboratories and into diverse real physical worlds, they always face a data conversion challenge—the physical world itself is continuous, unordered, and cannot be directly handed over to models for training.
While most teams in the industry are still accumulating data linearly by adding camera positions and expanding collector staff, InSpatio offers another solution: treating physical space as raw material and building an all-weather running embodied Token factory using 3D reconstruction, generative models, and parallel simulation. When millions of hours of physical interaction data begin to be scheduled monthly like industrial products, large-scale training for embodied intelligence finally has a stable data foundation.