EN / 中文

Huawei Cloud Releases CloudRobo-WM1.2, Ascend-Native Embodied World Model

by zhidongxi·November 17, 2026

Author | Sanbei, Editor | Xinyuan

Zhidx reported on November 16 that today, Huawei Cloud launched its first Ascend-native embodied world model, CloudRobo-WM1.2, and announced the official release of WM1.2 model weights, post-training inference code, and pre-training best practices on November 30. The model employs a shared world model backbone network to achieve prediction-enhanced world representation, and utilizes a world-model-native DualAct architecture for native action generation. It ranked first in both RoboTwin 2.0 and LiberoPro 2.0 evaluations.

At the event, Huawei Cloud demonstrated robots based on CloudRobo-WM1.2 completing real industrial scenarios such as continuous industrial-grade material handling and bearing burr cleaning, as well as storage tasks like stocking shelves and placing gifts into cabinets in an unmanned retail scenario.

During the conference, Huawei Cloud revealed: "Huawei Ascend Cloud leads the embodied AI market landscape in China, with 80% of Chinese embodied AI enterprises choosing Ascend Cloud, and near-complete coverage among key universities." The term "leading the landscape" likely refers to the highest ecosystem penetration rate and customer coverage within China according to the company's own statistical metrics, rather than third-party revenue-based rankings. Additionally, Huawei Cloud announced that intelligent sweeping robots, jointly developed with leading sanitation equipment manufacturer Fulongma Group, have started pre-sales.

01.Addressing Five Major Challenges: Huawei Uses Five Technologies to "Boost" Robot Intelligence

Wu Xiaodong, President of Huawei Cloud's Embodied Product Line and Director of the Embodied Intelligence Lab, disclosed that his team has been attempting to implement solutions in large industry models, autonomous driving for commercial vehicles, and the embodied intelligence sector, encountering numerous challenges. For instance, traditional visual representations become "half-blind" when occluded, "confused" when the perspective changes, and "dumb" when moving, leading to a significant drop in success rates in complex open environments.

Overall, as embodied intelligence transitions from laboratories to the real physical world, it faces multiple challenges including difficulty in multimodal physical understanding, coupling difficulties between world cognition and action generation, scalability issues with heterogeneous data, difficulty in generalization across different robot embodiments, and high computational costs for real-time inference.

In response to these challenges, leveraging engineering practices in AI and intelligent driving, Huawei Cloud introduced the CloudRobo-WM1.2 world model and the CloudRobo development platform, aiming to build a "silicon-based fertile ground" for embodied intelligence. CloudRobo-WM1.2 is built around core embodied intelligence challenges with five key technical capabilities:

1. PPA (Perceive-Predict-Act) Multimodal World Representation Architecture: The model not only identifies objects but also predicts the outcomes of physical interactions. It integrates RGB video, semantic features, depth, and point clouds, introducing future multimodal predictive supervision. Wu Xiaodong noted that the goal of unified representation modeling is to enable robots to progress from "clearly seeing a single image" to "understanding a 4K movie," thereby better comprehending 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 Pre-training Framework: Similar to drawing a caricature of Andy Lau where capturing distinctive features suffices for recognition without full detail, CloudRobo-WM1.2 organizes supervision based on the effective information actually present in the data. This prevents valuable data from being discarded due to local missingness, unlocking the scaling potential of heterogeneous data. Issues such as missing action annotations, invalid dimensions, lack of depth, and incomplete camera views are all compatible. Wu Xiaodong stated that the dynamic supervision mechanism allows the model to "handle both coarse and fine grain data," jointly supervising data with missing action labels, invalid dimensions, or absent depth to maximize data value.

3. World Model Native DualAct (Dual-path Unified Expert Training) Architecture: "One brain that understands the world, two paths for action generation." Analogous to human reflexes bypassing the cerebral cortex (knee-jerk reflex) for fast, direct control, versus conditioned responses learned through experience (fear of tigers). The world model directly outputs high-precision actions for native robot control; based on the MOT architecture, action experts can also achieve direct and efficient action output. The dual-channel action architecture design allows developers to flexibly define the size of action experts, improving post-training efficiency.

4. Embodiment Semantic Fidelity Cross-Embodied Motion Modeling: This acts as a unified motion translator for various robot forms, including single-arm, dual-arm, and wheeled robots, unifying semantics despite multi-embodiment differences. Wu Xiaodong mentioned that simply concatenating dimensions for multi-embodiment systems would lose individual variations ("thousand faces for thousand people"). The team constructed a robot translator to jointly train 96-dimensional robot motions and 168-dimensional human motions using native semantics, sharing a motion language and reusing human motion experience.

5. Four Flexibly Composable Elastic Inference Modes: World/Expert Action, Joint/Action-only: Like ordering dishes, this offers flexibility ranging from lavish to economical, matching customer compute resources and inference latency needs. It supports two output targets—Joint (outputting future world + high-precision actions) and Action-only (outputting actions only)—and two output heads—World Action (prioritizing high precision) and Expert Action (prioritizing speed and efficiency)—forming four composable inference modes. Customers can combine these flexibly based on compute and latency requirements.

In summary, these five core technologies collectively support the continuous evolution of CloudRobo-WM1.2 in world modeling, action generation, scalable data learning, and engineering deployment, providing a technical foundation for embodied intelligence to transition from model research to application in complex physical scenarios.

02.Three-Layer Model + Three-Layer Cloud: Huawei Cloud Builds a "Full-Stack Foundation" for Embodied Intelligence

Digging deeper into the five technical features, the CloudRobo-WM1.2 technical architecture is divided into three layers:

Layer 1: Embodied-Native Multimodal PPA Model Architecture: Combines prediction-enhanced world understanding with native dual-path actions, unifying the modeling of world understanding and action generation.

Layer 2: Embodied Data Scaling System: Features elastic pre-training with effective supervision and cross-embodied motion modeling with embodiment semantic fidelity, addressing the utilization of incomplete data and coordination of heterogeneous embodiments.

Layer 3: World Model Engineering Inference: Provides four-mode elastic inference, adapting to diverse scenario needs for prediction and action generation while optimizing computational efficiency.

Simultaneously, Huawei Cloud has built a cloud-native, full-stack end-to-end Ascend computing platform for embodied technology, structured in three layers:

Compute Infrastructure: An heterogeneous computing architecture tailored for multimodal and physics simulation, featuring fused heterogeneous compute and supernode interconnects. It uses techniques like context parallelism, parameter sharding, and activation recomputation to handle long-sequence video training.

Compute, Network, Data, Storage: Focuses on coordinated optimization of data flow and compute flow for embodied tasks. It enables unified PB-scale multimodal data storage and retrieval, joint scheduling of computation and communication for embodied RL, and improved stability of inference latency.

AI Framework: Accelerates Ascend-native multimodal and embodied reinforcement learning. Support for Ascend NPUs in RLinf v0.4 has been significantly enhanced, achieving up to a 2.78x improvement in throughput.

Wu Xiaodong noted that the team collaborated with Professor Yu Chao's group at Tsinghua University to deploy the RLinf reinforcement learning framework on the platform, resulting in a 2-3x improvement in inference efficiency and model training throughput.

03.Six Engineering Capabilities Bridge the "Last Mile" of Embodied Intelligence Implementation

There remains a significant distance between the birth of a world model and its final implementation in the real world. Wu Xiaodong stated that Huawei Cloud has created an end-to-end toolchain, enabling developers to easily build their own models and applications on the platform.

From model inference to actual physical operations, CloudRobo distills six engineering capabilities to support the scaled deployment of embodied intelligence from laboratories to various industries:

1. Simulation-Enabled Industry Cold Start: High-fidelity physics simulation, extreme condition synthesis, and Real2Sim2Real transfer reduce the cost of real-machine data collection.

2. Cloud-Native Model Training: Ascend-native acceleration for multimodal training and reinforcement learning, elastic compute scheduling, and a unified development toolchain for WAM/VLA/VLN.

3. NearLink & HarmonyOS Ecosystem Integration: High-concurrency R2C gateway supports keep-alive connections for 1K+ embodiments, with minute-level hot updates for GB-scale models.

4. R2C Cross-Embodiment Rapid Adaptation: Hour-level integration, unified multi-embodiment data and action interfaces, enabling model capability reuse across different robots.

5. Continuous Learning Evolution via RSI Architecture: Closes the loop by connecting operational data feedback, reward evaluation, real-machine reinforcement learning, and policy iteration.

6. Safe and Reliable Operation: Risk perception, autonomous recovery, task orchestration, and edge-cloud collaboration ensure the safety baseline for embodied intelligence.

Wu Xiaodong mentioned that Huawei Cloud is building a trinity strategy around embodied intelligence: "deepening technology, opening empowerment, and co-building ecosystems." This involves sustained investment in core technologies to drive model iterations, fully opening models and toolchains based on Ascend Cloud compute to empower industry partners, and aggregating ecosystem strengths to co-build industry solutions, accelerating the scaled deployment of embodied intelligence across industries.

04.Conclusion: World Model Competition Heats Up as Robots Learn to Think and Act

Globally, embodied world models are evolving from simple environment prediction to integrated prediction and action generation. Different technical approaches are all striving to answer how robots can truly understand physical laws and translate that understanding into executable actions. Huawei Cloud's differentiated layout in this instance lies in integrating world models with Ascend compute, data toolchains, simulation training, and robot deployment platforms, attracting industry partners through open weights.

Going forward, industry competition will focus not just on leaderboard rankings, but on cross-embodiment generalization, success rates in real-world scenarios, and deployment costs. Whoever first closes the loop from model capability to scaled delivery will truly hold the initiative for industrialization.