Huawei Cloud has released and open-sourced CloudRobo-WM1.2, an embodied world model built natively on its Ascend computing platform. The model has claimed first place on two evaluation benchmarks—RoboTwin 2.0 and LiberoPro 2.0—achieving average success rates of 88.98% and 81.7%, respectively. It enables robots to understand and associate environmental cues, predict changes, and generate actions. Trained entirely on domestic Ascend computing power, the model adopts a shared world-model backbone and a native DualAct architecture for action generation. Its pre-training framework is designed to handle incomplete training data with missing annotations and partial information, and it features cross-embodiment motion modeling capabilities—allowing it to adapt to robots of different forms. Developers can also select inference modes on demand to balance performance against operating costs. The open-source release includes model weights, inference code, the training framework, and best practices for Ascend computing power. The CloudRobo platform currently hosts more than 16,500 developers. Huawei Cloud said it will continue to roll out new models in the first quarter of next year and will offer up to 1 million yuan in Ascend Cloud computing power support to approved projects, aiming to accelerate robot deployment in industrial, retail, and other scenarios.
