AnswersPublished16 sources
How InSpatio Plans to Scale Robot Training Data With Simulation
InSpatio says its Hangzhou center aims to generate one million hours of robot training data per month by reusing editable 3D environments.
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- InSpatio says its Hangzhou center aims to generate one million hours of robot training data per month by reusing editable 3D environments.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: How does InSpatio, the spatial intelligence foundation model company spun out of SenseTime in 2025, aim to address robot training data short. Article summary: InSpatio’s approach is to make robot-training environments editable, reusable and scalable, rather than increase data collection solely by adding physical robots and training grounds; reporting confirms the Hangzhou cent. Topic tags: general, academic, education, general web, user generated. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, water
InSpatio’s plan is to make robot-training environments reusable and editable, so data variety can grow without rebuilding a physical training ground for every new layout. The company opened its simulation-led embodied AI data center in Hangzhou on October 29, 2026, and says it aims to produce one million hours of training data a month. That figure is an ambition, not independently demonstrated throughput. 4
25
The strategy: scale environments, not just demonstrations
Collecting robot data in physical spaces can involve repeated setup and resets. InSpatio’s reported alternative is to combine reconstructed or generated 3D scenes with simulation, then vary environments and tasks within them. The reporting describes a system intended to make scenes editable, reusable and scalable, with the goal of collecting not only more data but data from more spatial arrangements. 25
That distinction matters: a large number of demonstrations in a fixed layout may add volume without covering many different situations. Editable scenes could make it easier to change object placement and task conditions, although the available reporting does not quantify how much time or cost this approach saves compared with physical data collection.
The center’s reported production figures also need context. The coverage describes a target of one million hours per month, or about 100,000 hours per day. Those rates are not equivalent if sustained: 100,000 hours daily would amount to about three million hours over a 30-day month. The reports do not explain whether the daily figure is a peak, a separate target or a different accounting measure. 25
From real spaces to simulated scenes
One input route is real-world scanning. Hesai’s Kosmo system combines spatial capture hardware with processing intended to turn captured spaces into 3D outputs. Hesai announced the Kosmo 3D world data engine in October 2026. 23 Reporting on the system says a 100-square-meter indoor space can be scanned in about 15 minutes, with outputs including 3D Gaussian splats, meshes and depth data. These are reported capabilities, not independent measurements of reconstruction quality.
25
26
InSpatio’s Topos simulator is described as supporting 3D scene generation, real-world reconstruction and embodied simulation. 14 The company’s InSpatio-World research also describes recovering and generating dynamic interactive scenes from reference video.
17 Those sources support the broader scene-building approach, but do not independently confirm every input format or the full Topos Pro workflow described in some accounts.
A scan is not yet a robot-ready world
A reconstructed scene may look convincing without accurately representing how objects move or collide. Reporting on Kosmo notes that physical properties and other data still need to be completed or configured; scanning alone does not automatically create a fully specified simulation. 26
For robot training, this is a crucial step. Objects need to be represented in ways that allow meaningful interaction, and the environment must behave plausibly when a robot pushes, grasps or moves them. The available sources do not verify InSpatio’s exact procedures for assigning object properties, reconstructing hidden surfaces or calibrating contact physics. Those details should not be treated as established capabilities based on the production target alone.
Why task difficulty raises the bar
Robot training ranges from navigating around obstacles to manipulating objects and handling contact or articulated mechanisms. As tasks become more interaction-heavy, a scene’s visual appearance is less sufficient: geometry and simulated physical behavior both affect whether a training example is useful.
Simulation offers a potential way to repeat and vary scenarios without resetting a physical setup each time. But inaccurate physics can also produce misleading training data. The reporting available here does not verify a formal InSpatio L1–L3 delivery system, its real-robot acceptance tests or measured transfer results. Real-robot trials remain important for determining whether simulated training transfers to physical tasks.
What the announcement establishes—and what it doesn’t
The sources establish that InSpatio opened a simulation-led data center in Hangzhou and announced a million-hours-per-month production ambition. They also describe a broader pipeline linking spatial capture, 3D reconstruction and simulation. 4
25
They do not establish that the center has already reached that volume, that its data reliably cover the full range of robot tasks, or that policies trained there perform well on physical robots. The central promise is clear: create many more varied training scenarios by making 3D environments reusable. Whether the approach delivers the stated scale and real-world results remains to be demonstrated.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "How InSpatio Plans to Scale Robot Training Data With Simulation"?
InSpatio says its Hangzhou center aims to generate one million hours of robot training data per month by reusing editable 3D environments.