Guanglun Intelligence, Google, and NVIDIA Co‑Define Physical AI Simulation Standards

Over the past decade the AI bottleneck was compute power; the next decade the bottleneck will be data, and the prerequisite for generating that data is large‑scale, reproducible simulation. Without a scalable simulation world there can be no scalable robot data, and without a unified simulation standard there can be no true physical‑AI ecosystem.

Fei‑Fei Li is quoted as saying that bringing data into robot training is far harder than collecting images, because robots need massive amounts of interactive, executable, and verifiable physical interaction data. Such data cannot be scraped from the web or generated by raw compute alone, so simulation must generate it.

Traditional large‑model evaluation (training loss, benchmarks like MMLU or HumanEval) does not apply to physical AI, where a model’s performance can collapse with a change in lighting or surface material. Physical AI therefore requires a reproducible, parallelizable, and quantifiable evaluation pipeline, making simulation the prerequisite for both data production and capability assessment.

International giants are strategically positioning themselves in the simulation arena: NVIDIA acquired PhysX in 2008 and integrated it into Omniverse as the core of Isaac Sim; Google DeepMind bought MuJoCo in 2021, gaining control of the dominant robotics‑learning toolchain; Disney Research developed the Kamino engine for complex, closed‑chain mechanisms; other notable projects include Drake (MIT CSAIL → Toyota Research Institute) and Bullet, now tightly coupled with Google’s ecosystem. These moves aim to control how the world is modeled, how data is generated, and how robot capabilities are evaluated.

In September 2025 NVIDIA, Google DeepMind, and Disney Research will jointly open‑source the Newton beta engine, the first attempt to unify GPU‑native parallelism, high‑precision contact dynamics, complex mechanism solving, and robot‑learning ecosystems within a single open architecture.

Newton’s contributions are split among the partners: NVIDIA provides GPU‑native acceleration and the Warp framework; DeepMind contributes MuJoCo’s high‑precision dynamics migrated to GPU; Disney adds Kamino’s expertise in non‑standard configurations and extreme‑condition motion solving. Together they create a modular, GPU‑accelerated, auto‑differentiable simulation platform that serves as a unified foundation for physical‑AI training, evaluation, and deployment.

Guanglun Intelligence joins the Newton Technical Steering Committee as the sole Chinese core contributor, bringing its self‑developed “solve‑measure‑generate” stack, the SimReady asset system, and the high‑trust dynamics of Drake. This marks the first time a Chinese company participates at the core of global physical‑AI standard definition.

Beyond Newton, Guanglun has co‑released Isaac Lab‑Arena, the LeIsaac simulation platform (now listed in Hugging Face docs), and the RoboFinals benchmark, building a complete stack from low‑level engine to developer platform, evaluation framework, and industrial‑grade closed‑loop. The company positions itself not just as a data provider but as a leader shaping open‑source infrastructure, developer tools, and evaluation standards for the physical‑AI era.

The article concludes that simulation is transitioning from an auxiliary tool to the CUDA‑like standard layer for physical AI. Whoever defines the simulation standards—world representation, data generation, evaluation methods, and training pipelines—will shape the future of embodied intelligence, and the current window for setting those rules is rapidly closing.