Huawei Cloud launched CloudRobo-WM1.2 on 16 November, a robot world model trained on its Ascend chips, and says the model will be open source along with its training framework. TechNode flagged the plan the same day.
A world model predicts how a scene will change when a robot acts, and the robot plans its task from that. Wu Xiaodong, who runs Huawei Cloud’s embodied product line and its Embodied AI Lab, presented the model at an online event. Liu Liu, the lab’s chief model expert, explained how it’s built. Huawei Cloud calls it the first Ascend-native world model from its Industry AI Dream Factory, its programme for industry AI.
What Huawei says is inside
Huawei Cloud describes a shared world-model backbone that gives the robot a picture of its surroundings, sharpened by predicting what happens next. A second architecture, DualAct, turns that picture into actions. IT Home lists five features: the PPA multimodal world representation, DualAct, a pre-training framework that adapts to data labelled at different levels of detail, motion modelling that keeps a movement’s meaning across different robot bodies, and four inference modes that can be combined as needed.

The benchmark claims
Huawei Cloud says the model ranks first on two robot manipulation benchmarks, RoboTwin 2.0 and LIBERO-Pro 2.0. At the launch, slides gave its scores as 88.98% on RoboTwin 2.0 and 81.7% on LIBERO-Pro 2.0, according to Pandaily.
The RoboTwin 2.0 leaderboard confirms the first figure. CloudRobo-WM scores 95.28% there in the easy setting (clean scenes) and 82.68% in the hard one (randomised scenes), for an average of 88.98%. Awomo-0.6 comes next at 86.47% and GO-3 Nano Preview at 85.01%. Awomo-0.6 beats it on the hard setting alone, 84.78% to 82.68%. The entry was added on 15 November. An earlier one, CloudRobo-WM-Preview, went up on 19 October with an average of 81.95%.

On LIBERO-Pro 2.0, the benchmark’s own site doesn’t list CloudRobo-WM among its tested or upcoming models. The 81.7% rests on Huawei’s slides.
What’s open, and what isn’t yet
Cailian Press reports that Huawei Cloud will open the model weights, inference and code. The company also said it will open its training framework and its best practices for running world action models on Ascend. IT Home and Communications World headline the model as already open-sourced. Pandaily reports that the launch materials gave no download location, and that no weights or code repository were listed.
The platform behind it
The model sits on CloudRobo, the embodied AI development platform Huawei Cloud launched in June 2025. It covers data, model training, simulation and real-robot testing, and Huawei Cloud leaves the robot bodies to partners. Communications World reports Huawei’s figures: more than 16,500 developers, over 230 companies and universities, and 1,500-plus industry assets. Huawei also says 80% of China’s embodied AI companies use Ascend Cloud.

Cailian Press reports that Wu said new models are due in the first quarter of 2027, and that Huawei Cloud will give each approved project up to RMB 1 million of Ascend compute. Partner Fulongma Group said a street-sweeping robot built with Huawei Cloud is open for presale, with a launch in Shanghai in December. Fulongma says the robot’s perception has a world model built in.
None of Cailian Press, IT Home or Pandaily gives a release date for the weights.
Discover more from Sulat AI & Tech News
Subscribe to get the latest posts sent to your email.




