---
type: "post"
id: "inspatio-opens-simulation-led-embodied-ai-data-collection-c987ac"
slug: "inspatio-opens-simulation-led-embodied-ai-data-collection-c987ac"
title: "InSpatio opens simulation-led embodied AI data collection center in Hangzhou, targeting a million hours of robot data per month"
description: "InSpatio says it opened a simulation-led embodied-AI data collection center in Hangzhou on October 29, targeting one million hours of robot-training data per month."
retrieval_nugget: "InSpatio says it opened a simulation-led embodied-AI data collection center in Hangzhou on October 29, targeting one million hours of robot-training data per month. The important claim is not simply volume: the company is trying to turn scanned real spaces into editable, reusable simulation assets that can generate and test many task variations."
published_at: "2026-10-31"
updated_at: "2026-10-31"
record_date: "2026-10-31"
date_kind: "published_at"
topics: []
entities: ["InSpatio","Topos Pro","Hesai","Kosmo","SIDO"]
related_patterns: []
novelty_label: "new_angle"
what_changed: "InSpatio’s Hangzhou center reframes simulation as a way to scale robot-data diversity and production, targeting one million hours monthly through configurable environments and parallel computation."
prior_refs: ["memory:acb8c3ec-7195-807c-af9f-47273a3080b2","memory:35d303ff-787a-8149-ae57-12ada290a7c9"]
source_url: "https://www.163.com/dy/article/LAH1FEFM0511DPVD.html"
source_urls: ["https://www.163.com/dy/article/LAH1FEFM0511DPVD.html"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:inspatio-opens-simulation-led-embodied-ai-data-collection-c987ac"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/inspatio-opens-simulation-led-embodied-ai-data-collection-c987ac/","markdown":"https://newruntime.com/posts/inspatio-opens-simulation-led-embodied-ai-data-collection-c987ac.md","json":"https://newruntime.com/posts/inspatio-opens-simulation-led-embodied-ai-data-collection-c987ac.json"}
---

# InSpatio opens simulation-led embodied AI data collection center in Hangzhou, targeting a million hours of robot data per month

## Retrieval answer

InSpatio says it opened a simulation-led embodied-AI data collection center in Hangzhou on October 29, targeting one million hours of robot-training data per month. The important claim is not simply volume: the company is trying to turn scanned real spaces into editable, reusable simulation assets that can generate and test many task variations.

## Summary
InSpatio says it opened a simulation-led embodied-AI data collection center in Hangzhou on October 29, targeting one million hours of robot-training data per month. The important claim is not simply volume: the company is trying to turn scanned real spaces into editable, reusable simulation assets that can generate and test many task variations.

## What happened
According to AI Tech Review, spatial-intelligence company InSpatio opened what it describes as China’s first simulation-led data-collection center for embodied AI in Hangzhou on October 29. The company’s stated target is production at the scale of one million hours of training data per month, or 100,000 hours per day. Those are company targets rather than independently reported output figures.

The facility is built around a different production model from collecting demonstrations in physical training sites. The source describes ego-view collection and teleoperation as common routes for obtaining robot data, but both require people, equipment, and a configured physical environment. Changing a shelf layout, object placement, occlusion, or approach path can require rebuilding the setting and organizing another collection run. InSpatio’s proposal is to create configurable digital environments, where those conditions can be edited and recombined without repeating the full physical setup.

Its pipeline begins with real-world reconstruction. InSpatio says it works with Hesai’s Kosmo system to scan physical spaces; the article states that a typical 100-square-metre indoor space can be scanned in roughly 15 minutes and reconstructed with millimetre-level 3D Gaussian Splatting. InSpatio then uses its Topos Pro world simulator to convert inputs including single images, multiple images, video, LiDAR, and drone imagery into object-level scenes. The claimed result is a scene whose tables, shelves, objects, and other components can be selected, moved, replaced, and assigned physical attributes such as collision, mobility, and contact properties.

From a small number of real-robot or teleoperated demonstrations, the company says it combines motion planning, policy-driven generation, and automated physics checks to synthesize trajectories in hundreds of parallel environments. These trajectories are described as aligned across RGB images, depth, segmentation, 6D pose, and action labels. The source frames the capacity formula as environment configurations multiplied by parallel compute, rather than people multiplied by collection time.

InSpatio also describes three delivery levels for environments: L1 navigation environments with topology and traversable areas; L2 dexterous-grasping environments with millimetre-scale geometry, collision bodies, and dense occlusion; and L3 contact and articulated-dynamics environments for tasks such as opening drawers or insertion. It says simulation should extend and repeatedly test real-world data rather than replace final testing on physical robots.

## What the archive adds
This is a new event, but it belongs beside our note on [NVIDIA Cosmos 3 tools for simulation-specific models and physical-AI workflows](https://newruntime.com/posts/new-nvidias-cosmos-3-tools-simulation-specific-models-3af45b/). Both point to simulation assets and data pipelines becoming part of physical-AI development, rather than treating simulation as a standalone visualization step.

## Trend signal
No supplied Site Pattern is a sufficiently precise fit: this report concerns embodied-AI data production and simulation fidelity, not an established New Runtime pattern about agent software.

## Why now
The source places the launch in a setting where physical robot-data collection is constrained not only by raw hours but by the cost of introducing meaningful variation. InSpatio’s timing claim is therefore tied to a specific bottleneck: generating layouts, occlusions, object combinations, and task conditions faster than physical sites can be rebuilt.

## Connections to other events
The NVIDIA Cosmos 3 coverage provides the closest archive connection because it also concerns tools and pipelines for world-model and physical-AI work. InSpatio’s reported contribution is narrower and operational: a claimed center for reconstructing real spaces, editing them at object level, and producing simulated training trajectories at scale.

## Why it matters
For robot teams, the useful question is not whether simulated hours can be made cheaply, but whether a simulator can produce the failure cases that a policy actually needs. InSpatio’s proposed loop is concrete: identify a failure under a particular layout, occlusion, or contact condition; edit or generate environments around that condition; synthesize more trajectories; then validate the revised policy on real hardware.

That workflow could reduce dependence on repeatedly staging people, objects, and rooms for every variation. It could also make regression testing more systematic, provided the simulated geometry, contact properties, and action labels remain credible. The company’s L1-to-L3 framing is notable because it separates navigation-scale environments from the much harder requirements of grasping and contact dynamics.

## Limits and uncertainty
The supplied source is a report containing extensive company claims. It does not provide independently audited production figures, customer deployments, model-quality results, simulator-to-real transfer rates, or comparisons with other collection systems. A target of one million hours per month is not evidence that those hours are useful, diverse, physically faithful, or accepted by downstream robot teams. The reported scan time and reconstruction precision also do not establish performance for dynamic spaces, deformable objects, difficult lighting, or contact-heavy tasks. Most importantly, the source itself says final strategies still require validation on real robots.

## Skill to build now
Create a small failure-condition catalog for one robot task this week. Record the object arrangement, occlusion, approach direction, contact surface, and expected success criterion for each failure. Then turn the catalog into parameterized simulation cases and run the same policy against them after each change. This does not require accepting any vendor’s fidelity claims; it creates a test specification against which real and simulated runs can be compared.

## Reader takeaway
InSpatio’s announcement is a bet that robot-data scaling will come from reusable, editable environments rather than simply hiring more teleoperators or building more physical collection sites. The relevant metric to watch is not simulated hours alone, but whether teams can use those environments to reproduce failures and improve real-robot performance.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)