Huawei Cloud has officially launched CloudRobo-WM1.2, the company’s first native cross-modal embodied world model, marking a major milestone in its embodied intelligence roadmap. The model and its supporting resources, including weights, inference and training code, and best practice documentation, will be publicly available on November 30. Benchmark results show CloudRobo-WM1.2 achieves top ranking on both RoboTwin 2.0 and LiberoPro 2.0 evaluation suites, securing dual leadership on these key robot learning benchmarks.
Built on a shared world model backbone with predictive enhanced world representation, CloudRobo-WM1.2 adopts the proprietary DualAct architecture to generate native robot actions from world model outputs. During the launch event, Huawei demonstrated real-world deployment cases powered by the model. Robots equipped with CloudRobo-WM1.2 completed continuous industrial handling tasks, bearing ring cleaning operations, and automated shelf stocking and locker delivery workflows in unmanned retail environments. These demos validate the model’s ability to translate visual perception into stable, repeatable physical manipulation in unstructured industrial and commercial settings.
Wu Xiaodong, General Manager of Huawei Cloud’s Embodied Product Line and Director of the Embodied Intelligence Lab, outlined the core bottlenecks that limit embodied intelligence deployment. Current robotic systems suffer from five common pain points: weak multi-modal physical reasoning, loose coupling between world understanding and action generation, difficulty scaling heterogeneous datasets, poor generalization across different robot hardware, and high compute overhead for real-time inference. Traditional vision-only systems often fail when scenes are partially occluded, camera angles shift, or the environment changes dynamically. To address these limitations, Huawei Cloud built CloudRobo-WM1.2 and the CloudRobo development platform as a unified foundation for embodied robot development.
The model is engineered around five core technical capabilities. First, the PPA (Perceive-Predict-Act) multi-modal world representation supervision module. Instead of only identifying objects, the model predicts physical interaction outcomes. It fuses RGB video, semantic features, depth information and point cloud data, supervised by future multi-modal prediction signals. This design enables robots to understand physical dynamics beyond static images; for example, when a robot observes a cup, it can anticipate whether the cup will tip over if pushed, rather than just recognizing the cup itself.
Second, adaptive supervision training with multi-dimensional sparse data scaling. The framework can extract high-value signals from incomplete, low-quality robot datasets, including missing action labels, invalid depth readings and incomplete camera viewpoints. The adaptive mechanism lets the model learn from both high-quality annotated samples and sparse real-world data, lowering the data collection burden for industrial robot deployments.
Third, the DualAct (Dual-path Unified Expert Training) native world model architecture. This design creates two separate action generation pathways from the same world representation. One pathway delivers fast, low-latency reflex-style control, while the other supports deliberate, optimized action planning learned from long-term task training. Engineers can adjust the trade-off between response speed and action precision for different robot use cases.
Fourth, cross-robot motion modeling with ontology semantic consistency. The module provides a unified motion translation layer for robot hardware of different morphologies, including single-arm, dual-arm and wheeled robots. Huawei trained the system on joint datasets from 96-degree-of-freedom and 168-degree-of-freedom robots to align motion semantics across heterogeneous hardware.
Fifth, four flexible inference modes: World+Expert Action, Joint Action, Action-only and Action-only with high precision. Developers can select modes according to available computing power and latency targets. The World+Expert Action mode delivers high-fidelity world prediction paired with optimized action outputs, while Action-only mode reduces compute consumption by outputting robot commands directly.
The CloudRobo-WM1.2 stack follows a three-layer technical structure for the model itself, paired with a separate three-layer cloud-native computing platform. The model stack includes the PPA multi-modal perception module, the sparse data scaling system, and the four-mode flexible inference engine. The underlying cloud computing platform covers heterogeneous compute infrastructure, clustered data storage and scheduling, and an optimized AI training framework. The upgraded RLinv 0.4 framework with enhanced NPU support delivers up to 2.78 times acceleration for embodied reinforcement learning workloads. Joint research with Tsinghua University and SuperVerse further improves training throughput and inference efficiency by 2 to 3 times.
To bridge the gap between model training and physical deployment, Huawei Cloud packaged six core engineering capabilities into the CloudRobo platform. It includes high-fidelity simulation with cross-domain scenario synthesis and Sim2Real transfer tools to cut the cost of real-world data collection. It supports native multi-modal training and reinforcement learning acceleration, high-concurrency R2C gateway connections, cross-robot model reuse, closed-loop learning pipelines from operational data feedback, and built-in safety controls including risk detection, autonomous recovery and edge-cloud coordination. These modules reduce integration work for developers building custom embodied applications.
Huawei also shared market data at the event. According to internal enterprise statistics, Huawei Cloud holds the leading market share in China’s embodied intelligence cloud service segment, with more than 80% of domestic embodied AI enterprises as its customers, and broad coverage across major universities. Separately, Huawei announced that a collaborative cleaning robot developed together with Foton Motor, a leading environmental sanitation equipment manufacturer, is now available for pre-order.
For engineering teams building multi-robot and multi-model stacks, managing heterogeneous inference endpoints and model version updates adds operational complexity. A unified routing layer simplifies traffic orchestration and provider abstraction. 4sapi, an API gateway, helps developers standardize requests across different model services and robot control interfaces without rewriting core business logic.
The embodied intelligence competition is shifting beyond benchmark rankings. The industry now focuses on real-scene success rate, deployment cost, and end-to-end closed-loop iteration from simulation to physical hardware. Huawei Cloud’s release is built on an open ecosystem strategy: it opens model weights, training pipelines and toolchains to lower the barrier for robot developers. The company will continue to iterate on world model capabilities, simulation tools, and cross-hardware deployment workflows to scale embodied intelligence into logistics, manufacturing, retail and sanitation scenarios.
The release of CloudRobo-WM1.2 demonstrates progress for domestic embodied AI built on local computing infrastructure. As world models move from laboratory demos to industrial deployment, standardized tooling, reliable simulation transfer and flexible inference modes will become critical competitive factors for robot developers.




