Huawei Cloud on the 16th released CloudRobo-WM1.2, an embodied world model for robots optimized for its Ascend AI chips, and made it open source. The move is seen as an effort to bring robot AI out of the lab and into industry.
Huawei Cloud released and open-sourced CloudRobo-WM1.2, an embodied world model built on its Ascend platform, through its "Industry AI Dream Factory" on the 16th, Science and Technology Daily reported. A Xinhua-affiliated financial outlet's morning briefing on the 17th said Huawei Cloud plans to open the model's weights and inference code.
광고 문의 · 300×250Embodied AI, which lets robots understand the physical world and act in it, has become a key area of competition in China's tech industry. Science and Technology Daily noted that moving the technology from the lab into the real world has faced hurdles including multimodal physical understanding, coupling world cognition with action generation, scaling the use of heterogeneous data, generalizing across different robot bodies and high real-time inference costs.
Huawei Cloud said the model uses a shared world-model backbone to build prediction-enhanced representations of the world and generates actions through a world-model-native architecture called DualAct. Other core technologies include adaptive pretraining for multi-granularity supervised data, cross-embodiment motion modeling and four inference modes selectable on demand. The company said the model ranked first on the RoboTwin 2.0 and LiberoPro 2.0 robot manipulation benchmarks, though no scores were disclosed.
On the engineering side, the model combines Ascend Cloud's heterogeneous computing, unified storage and search for petabyte-scale multimodal data and native AI framework acceleration. Huawei Cloud said it supports the full embodied AI development cycle, from simulation-based cold start and cloud-native training to cross-embodiment adaptation and continual learning.
Huawei Cloud outlined three priorities, deep investment in core technology, open enablement and joint ecosystem building, and said it will open the model and toolchain to industry partners. The announcement did not include model size, training data volume or license terms. As Chinese firms accelerate robot AI development on domestic chip ecosystems, attention is turning to whether open-sourcing will translate into wider industrial use.