Independent Source For China's Tech Innovation Since 2000

Huawei Cloud Launches 'Industry AI Dream Factory' and Open-Sources Embodied World Model

Huawei Cloud has unveiled a new industrial development initiative dubbed the "Industry AI Dream Factory" alongside an open-source embodied world model designed to help robots better understand physical environments and perform complex real-world tasks, state financial newspaper Securities Daily reported Tuesday.

Built natively for Huawei's Ascend artificial intelligence computing architecture, the newly open-sourced model, named CloudRobo-WM1.2, uses a shared world model backbone network to generate prediction-enhanced representations of physical surroundings. It also incorporates a native DualAct architecture that allows the system to directly generate physical movement commands.

The release targets key technical hurdles that have slowed the transition of embodied AI from academic laboratories to commercial factories and service environments. Developers in the sector have struggled with teaching machines to comprehend dynamic physical interactions, linking environmental perception with motor control, scaling messy real-world training datasets, generalizing software across different robot body types and reducing real-time computing costs.

According to Huawei Cloud, CloudRobo-WM1.2 addresses those bottlenecks through upgrades in world modeling, action generation, large-scale data training and engineering deployment. Using a multimodal world representation architecture, the model moves beyond recognizing static images to predicting the physical outcomes of interactions, enabling robots to understand dynamic physical processes and execute native motion control.

To solve data scarcity and formatting inconsistencies, the model employs a multi-granularity supervised data adaptive pre-training framework. The system organizes supervised learning around whatever valid information is available within a given dataset, preventing valuable training material from being discarded due to missing action annotations, invalid dimensions, incomplete depth readings or partial camera angles.

The platform also tackles cross-embodiment generalization through semantic-preserving motion modeling. Acting as a unified "action translator," the shared model distinguishes the distinct motion mechanics of single-arm, dual-arm and wheeled robots, enabling collaborative learning across varied hardware designs.

Huawei Cloud said the release reflects a three-part strategy focused on core technical research, open platform empowerment and ecosystem development. Backed by Ascend cloud computing power and cloud-native infrastructure, CloudRobo-WM1.2 integrates heterogeneous computing, unified multimodal data storage and search, rapid indexing and native AI framework acceleration.

Those infrastructure tools provide end-to-end engineering support ranging from simulation cold starts and cloud-native training to cross-hardware adaptation, continuous learning and safe physical operation.

A Huawei Cloud representative said the company aims to build an open "silicon-based black soil," a foundational digital ecosystem uniting robot hardware manufacturers, AI model developers, universities, research institutes and enterprise clients to accelerate the large-scale industrial adoption of robotics.

Other China Tech Buzz: