Before robots can leave the lab, they first need to understand the world.

On the afternoon of November 16, Huawei Cloud held an online launch event for its CloudRobo embodied world model. Wu Xiaodong, president of Huawei Cloud's embodied intelligence product line and head of its Embodied Intelligence Lab, unveiled CloudRobo-WM1.2, the first Ascend-native embodied world model. He also announced that the model weights, inference implementation and code would be open-sourced, along with the training framework and best practices for WAM on Ascend computing power.

No. 1 on Both Benchmarks

The leaderboards offered an initial indication of the model's performance.

Trained on China's homegrown Ascend computing power, CloudRobo-WM1.2 uses a shared world-model backbone to create prediction-enhanced world representations and a world-model-native DualAct architecture for action generation. The model ranked first on both the RoboTwin 2.0 and LiberoPro 2.0 benchmarks, achieving average success rates of 88.98% and 81.7%, respectively.

To address common challenges such as multimodal physical understanding, cross-embodiment generalization and the high computing power cost of real-time inference, Huawei Cloud introduced five innovations: a PPA multimodal world-representation architecture, a world-model-native DualAct architecture, a multi-granularity supervised-data adaptive pretraining framework, embodiment-semantic-preserving cross-embodiment motion modeling, and four flexible inference modes that can be combined as needed: World Action, Expert Action, Joint and Action-only.

At the launch, Liu Liu, chief model expert at the Embodied Intelligence Lab, explained the model's architecture and capabilities and shared updates on robot deployments in industrial, retail and other settings.

Built on the CloudRobo Platform

The model did not emerge in a vacuum. It was built on the CloudRobo platform.

CloudRobo is Huawei Cloud's one-stop embodied intelligence development platform. Launched in June 2025 and underpinned by the Ascend Cloud, it covers the entire workflow from data to deployment. Its built-in Embodied Intelligence Plaza provides preloaded data, perception, planning and manipulation models, as well as simulation assets. The R2C SDK connects physical robots to cloud-based inference and control, while the underlying infrastructure supports flexible configurations of heterogeneous CPU, GPU and NPU computing power.

Wu said CloudRobo now has more than 16,500 developers, and Huawei Cloud will continue releasing new models in the first quarter of next year. On the ecosystem front, CloudRobo has brought together more than 230 companies and universities and accumulated over 1,500 industry assets spanning industrial, electric power, sanitation, healthcare and eldercare applications. Huawei Cloud will also provide up to 1 million yuan in Ascend Cloud computing power support for each approved project and work with partners to build high-quality industry datasets. According to information disclosed at the event, 80% of China's embodied intelligence companies have chosen the Ascend Cloud, which has also achieved near-complete coverage among the country's leading universities.

Beyond the Leaderboards, What Is It Trying to Solve?

Beyond its leaderboard results, the more important question is what the model is designed to solve.

At its core, a world model allows a robot to predict how a scene will change in response to its actions. The robot can first be trained and tested in a simulated environment before transferring those capabilities to physical hardware.

Huawei Cloud's objective is clear: enable models to understand the physical world and help embodied intelligence move from the lab into a broad range of industries. The company has again clarified its role in that strategy: it will provide computing power, models and a development platform to add intelligence to its partners' hardware, while leaving robot manufacturing to industry players.

From an institutional perspective, the market could be substantial. Huaan Securities describes physical AI as “the next major wave in AI,” with a potential market worth hundreds of billions to trillions of U.S. dollars. World models provide the underlying cognitive framework for physical AI, acting as an “AI brain” that understands physical laws, reasons about causal relationships and makes decisions. Embodied intelligence serves as physical AI's action layer. Together, the two form the foundational framework for cognition and action.

The Field Is Already Crowded

Huawei Cloud is far from the only company making a move. The field is already crowded.

NVIDIA is pursuing a full-stack platform strategy: the Cosmos world model understands and generates the physical world, the Isaac simulation platform handles training and validation, GR00T serves as a general-purpose brain for humanoid robots, and Jetson Thor supports on-device inference. At its GTC conference in March, NVIDIA named ABB, FANUC, KUKA, Yaskawa, Figure and AgiBot among its partners. Cosmos 3, released in early June, is billed as the world's first fully open-source, omni-modal model and can shorten physical AI training and evaluation cycles from months to days.

Tesla is concentrating its resources on robot hardware and mass production. Optimus has entered production on the manufacturing line at the company's Fremont, California factory, and Tesla plans to equip it with its internally developed AI5 chip around April next year.

In China, robot manufacturers including Unitree and AgiBot, as well as companies such as XPeng, Li Auto and BYD that are expanding into robotics from the automotive supply chain, are building out capabilities across robot models, data and hardware. Model developers including OpenAI, Google and Anthropic are also moving toward robotics through in-house research, external investments and technology development across multiple layers of the stack.

Huawei Cloud's Differentiator Lies in the Foundation

Each company is taking a different approach, and Huawei Cloud has its own strengths. Its key differentiator is its choice of foundation: the model runs entirely on China's homegrown Ascend computing power, while the weights, inference implementation, code and computing power best practices are all being open-sourced. This demonstrates the feasibility of training world models on domestic computing power and seeks to offer Chinese robot and model developers an independently controlled development path outside overseas platforms.