Huawei Cloud Releases Its First Ascend-Native Embodied World Model, CloudRobo-WM1.2
Huawei Cloud has launched its first Ascend-native embodied world model, CloudRobo-WM1.2, and announced that on November 30 it will open the model weights, post-training inference code, and pretraining best practices. The model adopts a shared world model backbone network and a native DualAct architecture, ranking first in both the RoboTwin 2.0 and LiberoPro 2.0 evaluations, and on-site demonstrations were given of real-world scenario experiments including industrial handling, bearing burr cleaning, and unmanned retail storage. Huawei Cloud also announced that an intelligent cleaning robot jointly developed with Fulongma Group has begun presales.
Why read this
SelectedAI score 78
Huawei Cloud's first Ascend-native embodied world model presents five technical designs and an open-weight timeline, which can be used to assess how it differs from existing approaches.
Topics
Body · AI translation
Just now, Huawei's first embodied world model has arrived! Trained on domestic computing power, it has taken two top spots. ZhiDongXi Download the app for exclusive early access. November 17, 2026, 08:43:27 From Beijing
ZhiDongXi reported on November 16 that today Huawei Cloud launched its first Ascend-native embodied world model, CloudRobo-WM1.2, and announced that on November 30 it will officially open the WM1.2 model weights, post-training inference code, and pretraining best practices.
The model uses a shared world model backbone network to achieve prediction-enhanced world representation, and is based on the world model's native DualAct architecture to achieve world-model-native action generation. It ranks first in both the RoboTwin 2.0 and LiberoPro 2.0 evaluations.
At the event, Huawei Cloud demonstrated experiments based on CloudRobo-WM1.2 in which robots completed continuous industrial-grade handling operations, bearing burr cleaning, and other real industrial scenarios, as well as storage tasks such as stocking items on shelves and placing gifts into lockers in an unmanned retail scenario.
At the meeting, Huawei Cloud revealed: "Huawei Ascend Cloud ranks first in China in the embodied intelligence market landscape; 80% of Chinese embodied intelligence companies have already chosen Ascend Cloud, and key universities are basically fully covered." The "first in landscape" here should refer to ecosystem penetration and customer coverage being first in China under the company's own statistical criteria, not revenue-based rankings from third-party institutions.
At the same time, Huawei Cloud announced that an intelligent sweeping robot jointly developed with the leading sanitation equipment company Fulongma Group has started presales.
1. Facing five major challenges, Huawei uses five technologies to "boost the brains" of robots
Wu Xiaodong, President of Huawei Cloud's Embodied Product Line and Director of the Embodied Intelligence Lab, revealed that his team has been trying to implement industry large models, autonomous driving for commercial and specialized vehicles, and the embodied intelligence industry, and has seen many challenges: for example, traditional visual representations become "half blind" once occluded, "confused" once the viewpoint changes, and "stupid" once things start moving, with success rates in complex open environments dropping noticeably.
Overall, embodied intelligence is moving from the laboratory into the real physical world and faces multiple challenges, including difficulties in multimodal physical understanding, the coupling of world cognition and action generation, the difficulty of utilizing heterogeneous data at scale, the difficulty of generalizing across robot embodiments, and the high cost of real-time inference computing power.
Facing these challenges, Huawei Cloud, based on its engineering practices in AI and intelligent driving, launched the CloudRobo-WM1.2 world model and the CloudRobo development platform, attempting to build a "silicon-based black soil" for embodied intelligence.
CloudRobo-WM1.2, centered on the core challenges of embodied intelligence, has built the following five core technical capabilities:
1. PPA (Perceive-Predict-Act) multimodal world representation architecture: the model not only recognizes objects but also predicts the outcomes of physical interactions. It integrates RGB video, semantic features, depth, and point clouds, and introduces future multimodal prediction supervision.
Wu Xiaodong said that the goal of unified representation modeling is to let robots go from "clearly seeing a single image" to "being able to understand a 4K movie," so as to better understand the physical world. For example, when a robot sees a cup, it not only knows what it is and where it is, but can also predict how it will move if pushed and whether it will tip over.
2. Multi-granularity supervised data adaptive pretraining framework: just as drawing a comic does not require drawing a complete Andy Lau, only reflecting his characteristics is enough for recognition. CloudRobo-WM1.2 organizes supervision according to the effective information actually available in the data, avoiding valuable data being discarded due to partial missingness, and unleashing the scaling potential of heterogeneous data. Data with missing action annotations, invalid dimensions in some parts, missing depth, and incomplete camera viewpoints can all be accommodated.
Wu Xiaodong said that the dynamic supervision mechanism should let the model "eat both coarse grain and fine grain," jointly supervising data with missing action annotations, invalid dimensions, and missing depth to maximize data value.
3. World-model-native DualAct (Dual-path Unified Expert Training) architecture: "the same brain that understands the world, two paths for action generation." This is similar to how humans have direct control that does not pass through the cerebral cortex (the knee-jerk reflex), which is fast, efficient, and direct, as well as conditioned reflexes established in life after learning (turning pale at the mention of a tiger). The world model directly outputs high-precision actions to achieve native robot action control; based on the MOT architecture, action experts can also achieve direct and efficient action output. The dual-channel action architecture design can support developers in flexibly defining the size of action experts and improving post-training efficiency.
4. Embodiment-semantic-preserving cross-embodiment motion modeling: equivalent to providing a unified action translator for robots of various forms such as single-arm, dual-arm, and wheeled robots, so that differences among multiple embodiments can be unified in semantics.
Wu Xiaodong mentioned that multiple embodiments cannot simply concatenate dimensions, otherwise the "everyone is different" distinctions would be lost; the team built a robot translator that jointly trains the native semantics of 96-dimensional robot motion and 168-dimensional human motion, sharing a motion language and reusing human motion experience.
5. Four flexible reasoning modes—World/Expert Action and Joint/Action-only—that can be combined on demand: like ordering à la carte, you can choose rich or lean according to your needs, matching customers' compute resources and inference latency on demand. It supports two types of output targets, Joint (outputting the future world + high-precision actions) and Action-only (outputting only actions), as well as two types of output heads, World Action (emphasizing high-precision output) and Expert Action (emphasizing speed and efficiency), forming four combinable reasoning modes. Customers can flexibly combine them according to compute and latency requirements.
Overall, the five core technologies jointly support the continuous evolution of CloudRobo-WM1.2 in world modeling, action generation, data-scale learning, and engineering deployment, providing a technical foundation for embodied intelligence to move from model research to application in complex physical scenarios.
II. Three model layers + three cloud layers: Huawei Cloud builds a "full-stack foundation" for embodied intelligence
Digging further along the five major technical features, the CloudRobo-WM1.2 technical architecture is divided into three layers:
First layer, embodied-native multimodal PPA model architecture: prediction-enhanced world understanding + native dual-path actions, unified modeling of world understanding and action generation.
Second layer, embodied data Scaling system: effective supervision elastic pre-training + embodiment-semantic-preserving cross-embodiment motion modeling, solving the utilization of incomplete data and collaboration among heterogeneous embodiments.
Third layer, world model engineering inference: four-mode elastic inference, adapting to the multi-scenario requirements of prediction and action generation as well as computational efficiency.
At the same time, Huawei Cloud has built a cloud-native full-stack embodied technology end-to-end Ascend cloud computing platform, divided into the following three layers:
Computing infrastructure: a heterogeneous computing architecture for multimodality and physical simulation, heterogeneous compute fusion, supernode interconnection, and technologies such as context parallelism, parameter sharding, and activation recomputation to handle long-sequence video training.
Computing-network-data-storage: collaborative optimization of data flow and compute flow around embodied tasks, PB-level unified storage and retrieval of multimodal data, joint scheduling of computing and communication for embodied RL, and improved stability of inference latency.
AI framework: Ascend-native multimodal and embodied reinforcement learning acceleration, with RLinf v0.4 Ascend NPU support greatly enhanced, achieving up to 2.78x throughput improvement.
Wu Xiaodong said that the team, together with Yu Chao's team at Tsinghua University, put the RLinf reinforcement learning framework onto the platform, improving inference efficiency and model training throughput by 2 to 3 times.
III. Six major engineering capabilities, bridging the "last mile" of embodied intelligence deployment
There is still a long way to go from the birth of a world model to its final deployment in the real world. Wu Xiaodong said that Huawei Cloud has built an end-to-end toolchain so that developers can easily develop their own models and applications on the platform.
From model inference to real physical operations, CloudRobo distills six major engineering capabilities to support embodied intelligence in moving from the laboratory to large-scale deployment across thousands of industries:
1. Simulation enables industry cold start: high-fidelity physical simulation, extreme-condition synthesis, Real2Sim2Real transfer, reducing the cost of real-machine data collection.
2. Cloud-native model training: Ascend multimodal training and native reinforcement learning acceleration, elastic compute scheduling, and a unified WAM/VLA/VLN development toolchain.
3. SparkLink and HarmonyOS ecosystem integration: a high-concurrency R2C gateway supporting 1K+ embodiments with concurrent keep-alive, and minute-level hot updates of GB-level models.
4. R2C cross-embodiment rapid adaptation: hour-level access, unified multi-embodiment data and action interfaces, and reuse of model capabilities across robots.
5. Continuous learning and evolution RSI architecture: a closed loop connecting operation data feedback, reward evaluation, real-machine reinforcement learning, and policy iteration.
6. Safe and reliable operation: risk perception, autonomous recovery, task orchestration, and device-edge-cloud collaboration, safeguarding the bottom line of embodied intelligence safety.
Wu Xiaodong mentioned that Huawei Cloud has built a trinity strategy of "deep technical cultivation + open empowerment + ecosystem co-building" around embodied intelligence: continuously investing in core technologies to drive model iteration and upgrading, fully opening up models and toolchains based on Ascend Cloud computing power to empower industry partners, and pooling ecosystem advantages to jointly build industry solutions, accelerating the large-scale deployment of embodied intelligence across thousands of industries.
Conclusion: The world model race heats up, and robots begin to learn to think and act
Globally, embodied world models are moving from pure environment prediction toward the integration of prediction and action generation. Different technical routes are all answering how to enable robots to truly understand physical laws and translate that understanding into executable actions.
Huawei Cloud's differentiated layout this time lies in connecting the world model with Ascend computing power, data toolchains, simulation training, and robot deployment platforms, and attracting industry partners through open weights.
Next, industry competition will not only look at leaderboard rankings, but also at cross-embodiment generalization, success rates in real scenarios, and deployment costs. Whoever is the first to run through the closed loop from model capabilities to large-scale delivery will truly hold the initiative in industrialization.
"Special statement: The above content (including videos, pictures, or audio if any) is uploaded and published by users of "Dafeng Hao", a self-media platform under Phoenix News. This platform only provides information storage space services.
Notice: The content above (including the videos, pictures and audios if any) is uploaded and posted by the user of Dafeng Hao, which is a social media platform and merely provides information storage space services.”
AI Trends