As embodied intelligence (Embodied AI) moves from the lab to the real world, a critical bottleneck has emerged: how to scientifically, efficiently, and scalably evaluate whether a robot is truly "intelligent"? Recently, World Labs, an AI pioneer company founded by Fei-Fei Li, and Guanglun Intelligent, a leading enterprise in simulation technology, announced a deep collaboration to build the world's first high-fidelity, scalable evaluation system for embodied intelligence. This marks a new stage in the field, shifting from "demo-driven" to "evaluation-driven."
The core of this collaboration lies in closing the loop between the "virtual world" and "physical capabilities." World Labs' Marble platform can generate highly realistic 3D physical environments—ranging from living rooms to industrial warehouses—with rich details, precise lighting, and realistic gravity and material properties. Meanwhile, Guanglun Intelligent contributes its unique technology stack: a GPU-accelerated physics solver ensures the authenticity of action interactions, an automated virtual-real alignment system aligns simulation results with real robot behavior, and the SimReady asset generation tool quickly builds standardized test scenario libraries.
The joint evaluation framework allows developers to test a robot's perception, planning, and execution capabilities in thousands of virtual environments in parallel. For example, a household robot can repeatedly practice "finding soy sauce and handing it to a human" in hundreds of kitchen layouts generated by Marble. The system automatically records dozens of metrics such as success rate, path efficiency, and object recognition accuracy, generating quantifiable reports that can be compared horizontally.
This breakthrough holds significant importance. In the past, progress in embodied intelligence relied mostly on single-scenario demonstrations or small-scale field tests, which were hard to reproduce, costly, and lacked a unified standard. Now, through high-fidelity simulation combined with automated evaluation, development cycles can be shortened several times over, and algorithm iteration efficiency can be greatly improved. More importantly, it provides an "equal playing field" for the industry—startups no longer need to build expensive testing facilities themselves, but can verify model performance in standardized environments.
Guanglun Intelligent's technology ecosystem plays the role of "infrastructure" in this process. Its GPU-based physics engine supports large-scale concurrent simulation, the SimReady asset library covers mainstream scenarios such as home, logistics, and manufacturing, and its automated alignment capabilities ensure that skills learned in simulations can be effectively transferred to physical robots.
Fei-Fei Li has repeatedly emphasized: "Spatial intelligence is the next frontier of general artificial intelligence." This collaboration is a practical implementation of that concept. When robots are no longer relying solely on one impressive demonstration to gain attention, but instead prove their reliability through repeatable and quantifiable evaluations, the commercialization path of embodied intelligence will finally have a solid foundation. This evaluation revolution driven by simulation may be the most crucial step before robots enter households around the world.
