Simulation and Data Platform Ecosystem Integration Research

1. Introduction

This piece is the capstone integration article of the 07-Embodied AI Group. Its subject of study is not a single vendor, but the infrastructure layer of Embodied AI: simulation platforms, world models, and datasets / data-collection infrastructure. Every vendor's model iteration in the preceding nine articles — whether 宇树's WVLA, Figure's Helix, or PI's π series — relies on this infrastructure layer to supply training data and evaluation environments. The main axis of differentiation within the group is the "infrastructure layer": whoever controls simulation and data controls the iteration speed of Embodied AI.

1.1. Note on the Integration Article: The Infrastructure Layer

Simulation and data platforms span three layers in the Harness Six-Layer Model: synthetic data generation serves L1 (Context / Training Corpus) and L4 (Data Assets); simulation evaluation is directly the embodied form of L5 (Evaluation and Observation); the licensing system and industry standards belong to L6 (Governance). This article unfolds in four segments — "International Stack — Competing Engines — Domestic Stack — Data Foundation" — and provides a closed-loop view of simulation, data, and training.

1.2. Three-Layer Ecosystem Structure

2. Glossary

3. Feature Description and Ecosystem Landscape

3.1. International Stack: NVIDIA as the De Facto Standard

The NVIDIA Isaac platform is the current de facto standard for robot simulation; its official positioning is an "AI Robot Development Platform," composed of simulation (Isaac Sim), robot learning (Isaac Lab), CUDA-accelerated libraries, AI models, and reference workflows. Key capabilities:

1. Isaac Sim: Omniverse + PhysX 5 + RTX rendering, supporting digital twins and synthetic perception data generation.
2. Isaac Lab: BSD-3 open-source RL framework, 4096–8192 parallel environments on a single GPU, with RSL-RL / RL-Games / SKRL / SB3 built in, and 16+ pre-configured robot models such as G1 / H1 / Spot / ANYmal / Franka; from 3.0 beta (2026-03) onward, Newton + MuJoCo Warp solvers are optional, and it can be used independently without the Isaac Sim runtime.
3. Cosmos World Model Series: five modules — Curator / Predict / Transfer / Reason / RL; Cosmos 3 was released at GTC 2026, positioned as a "Physical AI frontier foundation model"; the Physical AI Data Factory Blueprint unifies automated data generation.
4. GR00T + OSMO + Isaac Lab-Arena: the three-piece set of humanoid foundation model, compute orchestration, and policy benchmark platform.
5. Ecological position: at GTC 2026, NVIDIA named Physical AI one of its five strategic pillars; by early 2026, 110 robot "brain" developers and industry partners were already building on its stack.

Key licensing points (L6 material): Isaac Sim source code is Apache 2.0, but it depends on the Omniverse Kit SDK and assets, and redistribution requires an NVIDIA AI Enterprise license — a combination of "open-source code, controlled runtime."

3.2. Competing Physics Engines

3.3. Domestic Stack

Accompanying governance action: 工信部 released the 《人形机器人与具身智能标准体系（2026 版）》, the first standard system to cover the entire industry chain.

3.4. Datasets and Data-Collection Foundation

The two major open-source real-machine datasets constitute the public goods of the data layer: AgiBot World (智元, open-sourced 2024-12, million-scale real-machine data, 100+ tasks across five major scenarios, more than 10× the scale of DROID) and RoboMIND (北京人形机器人创新中心, open-sourced 2025-12, 300,000+ dual-arm manipulation trajectories, 700+ practical tasks). On the data-collection foundation side, 北京人形's data base (nearly 6,000 square meters, 40+ simulated real-scene zones, 150+ robots across 40 configurations + 100+ bodyless devices, "matrix-style collection") represents the highest-specification data-collection facility in China at present. Industry data baseline: the whole industry has about 500,000 hours of high-quality real physical interaction data, while a usable level requires about 10 million hours — a gap of about 20× (single review-article caliber, Grade C, citations must be annotated).

4. Platform Architecture

4.1. NVIDIA Full-Stack Layering

NVIDIA's embodied AI stack, from top to bottom, is: Cosmos World Model (data generation) → Isaac Sim / Lab (simulation and training) → GR00T (humanoid foundation model) → OSMO (orchestration) → Jetson Thor (on-device inference); the three hardware platforms DGX / OVX / AGX respectively support training, simulation testing, and deployment. Isaac Lab 3.0 introduces Newton + MuJoCo Warp as optional solvers and supports running independently without Isaac Sim, showing NVIDIA's rebalancing between "platform lock-in" and "open decoupling."

4.2. RoboMIND Data Super-Factory

北京人形机器人创新中心's "data super factory" connects the full closed loop of "collection → cleaning → quality inspection → annotation → training → real-machine evaluation → deployment → reflow" (24-hour automated quality inspection): annual production capacity 180,000 hours, nearly 30,000 hours already delivered, 70% serving industry customers, long-term target 1 million hours. One of its landmark results is that the 天工 robot, supported by the data of this base, ran the 100-meter dash in 8.64 seconds.

4.3. Simulation—Data—Training Closed-Loop Diagram

图 10-1｜仿真-数据-训练-真机部署的产业闭环（示意）

示意：基于本文分析绘制；数据缺口与 Sim2Real 数字为第三方口径并已在正文标注可信度等级。

5. Harness Design

The perspective of the Harness analysis in this article is special: simulation and data platforms themselves are upstream components of other vendors' Harnesses. The following gives observations on the "infrastructure layer" itself across the six layers.

5.1. L1 Context Engineering Layer

Synthetic data and world models are the "pre-training corpus factory" of the embodied field: Cosmos's automated data generation and the Physical AI Data Factory Blueprint are essentially mass-manufacturing scene priors (L1 context) within simulation. Open-source datasets such as RoboMIND / AgiBot World make the context of the real world a public good, lowering the L1 threshold for all downstream models.

5.2. L2 Tools and Execution Layer

Isaac Lab's 4096–8192 parallel environments on a single GPU make "tool execution" possible at tens-of-thousands scale concurrency; Genesis's 10,000+ parallel environments, and the GPU solvers of MuJoCo Warp / Newton, are all doing the same thing — pushing down the trial-and-error cost of physical execution to near zero. The real-machine part of the execution layer (Jetson Thor, Isaac TeleOp) extends the NVIDIA stack downward to the robot body.

5.3. L3 Orchestration and Control Layer

OSMO is the clearest infrastructure product at this layer: cloud-native workflow orchestration, chaining cross-cluster scheduling of data generation, training, evaluation, and deployment — equivalent to a CI/CD pipeline in the embodied AI field. On the domestic side, Genie Studio (智元) plays a similar role (see 02-agibot.md).

5.4. L4 Memory and State Layer

The data base and datasets are precisely the industry-level L4: RoboMIND's "matrix-style collection" (150+ robots across 40 configurations + 100+ bodyless devices) deposits experience in standardized structure; 觅蜂科技's 1 million hours of bodyless data shows that "experience assets decoupled from the robot body" are already scalable in production.

5.5. Evaluation and Observation Layer

This is the infrastructure layer's biggest contribution to the Harness system: Isaac Lab-Arena (general-purpose policy simulation evaluation) and the π0.5 academic baseline (see 08-physical-intelligence.md) constitute the embryo of a benchmarking public good; RoboMIND's "training → real-machine evaluation → deployment → reflow" closed loop turns L5 from a score into an operational process. Stanford HAI AI Index 2026 gives the core metric of this layer — simulation success rate 89.4% vs. real home 12%, a 77-percentage-point Sim2Real gap, showing that there is systematic distortion between the current L5 simulation evaluation and real-machine performance.

5.6. L6 Governance and Safety Layer

Three pieces of governance evidence: Isaac Sim's open-source–commercial boundary design of "source code Apache 2.0, runtime dependencies under controlled licensing"; 工信部's 《人形机器人与具身智能标准体系（2026 版）》 full-industry-chain standard framework; the CR-category compliance practice of the RoboMIND dataset and 24-hour automated quality inspection (data quality governance). Data sovereignty (attribution and cross-border transfer of real-machine scene data) is a governance blank where no standard has yet formed.

5.7. Six-Layer Strength Summary

6. Actual Cases

6.1. RoboMIND: Data Infrastructure with 20 Million Downloads

北京人形机器人创新中心's RoboMIND dataset, since being open-sourced in 2025-12, announced on 2026-09-02 that global downloads had broken 20 million (doubling within one month). Its data base is nearly 6,000 square meters, 40+ simulated real-scene zones, matrix-style collection with 150+ robots across 40 configurations plus 100+ bodyless devices, annual production capacity 180,000 hours, nearly 30,000 hours already delivered, 70% serving industry customers, long-term target 1 million hours; it connects the full closed loop of collection → cleaning → quality inspection → annotation → training → real-machine evaluation → deployment → reflow, with 24-hour automated quality inspection. Relying on the data of the base, the 天工 robot achieved 8.64 seconds in the 100-meter dash. This is the most complete public sample of the "data as infrastructure" narrative.

6.2. Quantitative Evidence of the Sim2Real Gap

Stanford HAI AI Index 2026: robot manipulation simulation success rate 89.4%, real home environment only 12% — the 77-percentage-point Sim2Real gap is the boundary condition that all "simulation evaluation" claims of this group must face head-on. Related to this, the industry's high-quality real-machine data is about 500,000 hours vs. demand of about 10 million hours (20× gap, Grade C single-source caliber), and institutional estimates put the 2026 domestic embodied data services market size at over 150 billion yuan. The two sets of numbers jointly explain why data collection factories (智元, 北京人形, 优必选自贡) appeared densely in 2026: the 77 percentage points that simulation cannot solve, in the short term can only be patched with real-machine data.

6.3. The Three-Sided Debate over Data-Collection Quality

On 2026-09-09, 新浪财经 / 第一财经 recorded representative divergences over the data-collection route; the three sides' views are presented in parallel:

This debate has no referee — the three routes (synthetic data / real-machine data collection / deployment reflow) are each heavily invested in by 银河通用, 智元, and Figure respectively, and there will be no conclusion before the end of 2026. For buyers and researchers, the correct posture is to make "data route" a must-ask item when evaluating any embodied AI vendor.

7. Summary

7.1. Strengths and Limitations

The "strengths and limitations" of this article correspond to the industry judgment of the infrastructure layer: the strengths lie in — NVIDIA's full stack has already polished the simulation-training-deployment toolchain to usable and replicable; the emergence of open-source datasets and benchmarking public goods has raised the industry's starting point as a whole; the domestic stack has formed local supply in the Sim2Real closed loop and data services. The limitations lie in — the 77-percentage-point Sim2Real gap shows that simulation evaluation is far from the level of "trusted substitute for the real machine"; the 20× real-machine data gap and the route have not converged; NVIDIA's open-source code but controlled-runtime licensing structure harbors supply chain dependence; data sovereignty and safety standards remain blank.

7.2. Applicability Boundaries

The subject of this article applies to all developers and buyers of embodied AI: algorithm teams choose the simulation stack (academia chooses MuJoCo / Genesis, industry chooses Isaac, LLM-driven automation chooses Genie Sim 3.0); data demanders choose datasets and data collection services (starting from the public goods RoboMIND / AgiBot World, procuring data collection services for specialized scenarios); strategic researchers take the "data route contest" as the main line of observation.

7.3. Selection Recommendations

1. For simulation stack selection, ask about licensing first: Isaac Sim's source code is open but asset redistribution is controlled; before commercial redistribution, confirm NVIDIA AI Enterprise authorization.
2. For starting datasets, use open-source public goods (RoboMIND / AgiBot World), and procure data collection services for the gap part; when procuring, price by verifiable metrics such as "95%+ pass rate class" rather than by hours.
3. When evaluating any embodied AI vendor, list "data route" as a must-ask item: whether to choose one of synthetic data, real-machine data collection, or deployment reflow, or to combine them, directly determines its generalization ceiling.
4. Pay attention to the progress of Newton (Linux Foundation governance) and 工信部's standard system: the dual-track evolution of physics engine governance and national standards will determine the toolchain landscape for the next three years.

Information Gap Statement

1. Isaac Sim 5.x / Isaac Lab 3.0 official release notes have not been obtained item by item; version details are cross-verified against developer pages and third-party archives.
2. Genesis official paper and repository main URL: the official site redirects to GitHub (Genesis-Embodied-AI); the paper link has not been directly verified.
3. NVIDIA Cosmos 3 official blog has not been directly fetched; Cosmos 3 information is based on Analytics Insight's retelling.
4. The data gap of "500,000 hours vs. 10 million hours" is the caliber of a single review article (Grade C); this article has annotated it, and it is not suitable for standalone citation.
5. Domestic data services market over 150 billion yuan is an institutional estimate caliber; there is no official statistics.
6. The number of deployed customers and revenue of each domestic simulation platform (RealMirror / SimOne, etc.) have not been disclosed.

8. References

1. NVIDIA Isaac — AI Robot Development Platform — NVIDIA Developer, 2026. https://developer.nvidia.com/isaac
2. NVIDIA Isaac Lab — NVIDIA Developer, 2026. https://developer.nvidia.com/isaac/lab
3. Newton, the Next-Generation Open-Source Physics Simulation Engine — NVIDIA Developer, 2026 (Beta repository: https://github.com/newton-physics).
4. Best Physical AI Development Tools and Frameworks in 2026 — Analytics Insight, 2026. https://www.analyticsinsight.net/amp/story/artificial-intelligence/best-physical-ai-development-tools-and-frameworks-in-2026
5. Robot SDKs, ROS 2, and Simulation Environments — Kentino, 2026. https://kentino.com/ja/blogs/build-guide/robot-sdks-ros-2-and-simulation-environments
6. NVIDIA Isaac Lab — AI Wiki (software stack / licensing / hardware requirements), 2026. https://aiwiki.ai/wiki/isaac_lab
7. Genesis open-source repository — Genesis-Embodied-AI / GitHub, 2026. https://github.com/Genesis-Embodied-AI/genesis-world
8. Beijing's "Data Super-Factory" Helps Robots Understand the World (RoboMIND 20 million downloads) — 人民网北京频道, 2026-09-03. https://bj.people.com.cn/BIG5/n2/2026/0903/c14540-41685467.html
9. Over 20 million downloads! 北京人形 accelerates the building of its "Data Super-Factory" — 搜狐, 2026-09-02. https://m.sohu.com/a/1070996577_163278
10. The "Power of Data" Behind the 天工 Robot's 8.64 Seconds — 腾讯新闻, 2026-09-02. https://news.qq.com/rain/a/20260902A0B17100
11. Embodied AI Data Soars, but Full Model Verification Has Not Yet Been Achieved — 新浪财经 / 第一财经, 2026-09-09. https://k.sina.com.cn/article_1733360754_6750fc7202001in52.html
12. The "Power of Data" Behind 天工's 8.64 Seconds (Five Key Points of the Data Base) — 具身之家, 2026-09. http://www.jushenhome.com/news_d.php?id=1650
13. Beste Roboter-Simulatoren für RL 2026 — Silicon Valley Robotics Center, 2026. https://www.roboticscenter.ai/de/rl-environments/best-2026