Huawei Cloud open-sources CloudRobo-WM1.2, an embodied world model that outputs actions directly and tops two robot benchmarks
Synopsis
Huawei Cloud released and open-sourced CloudRobo-WM1.2 in Hangzhou, an Ascend-native embodied world model, aiming to move robots from recognizing static images to anticipating physical interactions and generating actions directly. Embodied intelligence has been held back by difficult multimodal physical understanding, tight coupling between world cognition and action generation, hard-to-scale heterogeneous data, weak cross-embodiment generalization and high real-time inference cost. The model uses a shared world-model backbone for prediction-enhanced world representation and a world-model-native DualAct architecture to output actions, ranking first on both RoboTwin 2.0 and LiberoPro 2.0.
Interpretation
Huawei Cloud released and open-sourced CloudRobo-WM1.2, an Ascend-native embodied world model in which robot control no longer rests only on recognizing static images: the world model anticipates the outcome of physical interactions and emits actions directly. Moving embodied intelligence from the lab into the physical world runs into difficult multimodal physical understanding, hard coupling between world cognition and action generation, heterogeneous data that is hard to use at scale, weak cross-embodiment generalization and high real-time inference cost. CloudRobo-WM1.2 ranks first on both the RoboTwin 2.0 and LiberoPro 2.0 benchmarks.
A single shared world-model backbone carries the prediction-enhanced world representation, and the world-model-native DualAct architecture outputs actions directly, giving native robot action control. DualAct is a dual-channel action architecture built on the MOT architecture; action experts can also output actions directly and efficiently, and developers can flexibly size the action experts to improve post-training efficiency. CloudRobo-WM1.2 is built from five core technologies: besides DualAct, they are PPA multimodal world representation, multi-granularity supervised adaptive pretraining, embodiment-semantics-preserving cross-embodiment motion modeling, and four composable elastic inference modes spanning World/Expert Action and Joint/Action-only.
On the training side, supervision is organized by the effective information each datum actually carries, so data with missing action labels, partially invalid dimensions, missing depth or incomplete camera views can still be used, unlocking the scaling potential of heterogeneous data. Organizing supervision by each datum's actual effective information avoids discarding valuable data wholesale because of local gaps. The cross-embodiment component gives single-arm, dual-arm and wheeled robots a unified action translator, letting the shared model distinguish each robot's motion language and supporting multi-embodiment collaborative learning.
Perspective
The demonstrated scope is the model release and open-sourcing, plus first place on the RoboTwin 2.0 and LiberoPro 2.0 benchmarks; the intended readers are embodied-model developers and robot-body makers, and the setting relies on Huawei Cloud's Ascend Cloud compute and cloud-native infrastructure.
The text gives only the rankings on two benchmarks, without scores, comparison baselines or performance in real physical tasks; the practical effect of cross-embodiment generalization and the elastic inference modes is stated only as a design goal.
