As 2026 dawns, AI visionary Li Feifei's World Labs officially partners with simulation and synthetic data specialist Embody AI. This collaboration signals the AI industry's systematic move to tackle a pivotal bottleneck in embodied intelligence: large-scale evaluation.

While World Labs' world model clarifies the origin of the world, Embody AI's technology provides a means to measure progress. The synergy between these two entities transcends mere commercial strength; it heralds a new epoch where embodied intelligence is propelled by evaluation.

The realm of embodied intelligence is currently at a crossroads. Model capabilities are outpacing academic benchmarks at an alarming rate, yet there exists a dearth of precise and scalable methods for assessment. The industry's present state is marked by captivating demo videos, but the actual capabilities remain elusive in terms of quantification and comparison.

Yang Haibo, co-founder of Embody AI, identifies three major hurdles in the industry: academic benchmarks are lagging behind model advancements; real-world testing is both costly and time-intensive; and traditional simulation evaluations are confined to idealized scenarios.

Li Feifei foresaw this challenge years ago. During her tenure at Stanford, she championed the BEHAVIOR research series, aiming to establish a long-term evaluation framework for embodied intelligence akin to ImageNet. At the BEHAVIOR Challenge award ceremony during the 2025 NeurIPS conference, the presence of Embody AI CEO Xie Chen sent a clear message: evaluation is transitioning from a theoretical concern to an engineering and infrastructure imperative.

This transformation is underpinned by a fundamental shift in data paradigms. The China Academy of Information and Communications Technology highlights that while the debate over current robot data training paradigms lingers, synthetic and simulated data are garnering increasing attention.

Embody AI stands out in this landscape—according to industry veterans, over 80% of the simulation assets and synthetic data utilized by leading international embodied intelligence teams originate from Embody AI. As an industry infrastructure provider, it is uniquely positioned and qualified to assume the evaluation mantle.

The collaboration between World Labs and Embody AI is anchored in a core tenet: Digital Cousin. Proposed by Stanford SVL, this concept posits that the essence of simulation lies in its physical and spatial structure; as long as the structure is credible, details can be approximated.

World Labs' product, Marble, embodies this philosophy. As the world's inaugural commercial multimodal world model, Marble can craft high-fidelity 3D worlds from text, images, or videos.

Crucially, Marble generates persistent, downloadable 3D environments, mitigating issues of scene deformation and inconsistent details.

However, Marble adopts a visual world model approach and lacks physical parameters such as touch and gravity. To be genuinely useful for robot training and evaluation, it necessitates a physics engine and simulation assets that align with real-world physics.

This is precisely where Embody AI shines. It has constructed a tripartite simulation technology system for Marble, encompassing self-developed GPU physics solvers, a fully automated virtual-to-real physical measurement factory, and the capability to map real-world physical characteristics into SimReady assets and scenarios. Under this system, simulation transcends its role as a mere data generation tool and evolves into a repeatable and scalable data production system.

Within the embodied intelligence ecosystem, Embody AI has carved out its niche. Madison Huang, Senior Director of NVIDIA Omniverse and Physical AI, concedes, "Many internal NVIDIA projects rely on Embody AI's support."

In just two years, this company has amassed an impressive clientele that spans nearly all enterprises in the AI ecosystem requiring simulation and synthetic data: from model giants like NVIDIA and Google to robotics firms such as Figure AI and ZhiYuan Robotics, and industry behemoths like Toyota, Bosch, and BYD.

Towards the end of the previous year, Embody AI unveiled the RoboFinals evaluation standard, the industry's inaugural industrial-grade, scalable, and authentically reliable simulation evaluation platform. Concurrently, Embody AI collaborated with NVIDIA to craft Isaac Lab Arena, NVIDIA's new-generation open-source simulation evaluation framework. These endeavors are revolutionizing industry R&D processes. According to Yang Haibo, by integrating with simulation platforms, enterprise development cycles that once spanned 3 to 6 months can now be condensed to 2-3 weeks. The cost reductions and efficiency gains make rapid iteration of embodied intelligence feasible.

Embody AI's proposed simulation ecosystem concept aligns perfectly with this trend. Yang Haibo underscores, "Achieving simulation excellence demands not just technological breakthroughs but also the construction of a comprehensive simulation ecosystem. Without robust ecological support, simulation platforms cannot sustain development and iteration."

At the policy level, China's National Development and Reform Commission has explicitly proposed backing technological breakthroughs in the integration of simulation and real-world data, as well as promoting the construction of infrastructure such as training and pilot platforms. This indicates that the significance of simulation and evaluation as industry infrastructure has garnered national recognition.

http://kjt.gxzf.gov.cn/dtxx_59340/kjdt/t27113354.shtml

https://www.zidonghua.com.cn/news/newproduct/81150.html

https://finance.sina.com.cn/tech/csj/2026-01-19/doc-inhhvfrx8043654.shtml