---
title: "Huawei Cloud Open-Sources CloudRobo-WM1.2, a World Model for Robots Trained on Domestic Ascend Compute"
date: 2026-11-18
category: Foundation Models
site: NeuroAI
canonical: https://neuroai.site/a/na-model-huawei-cloudrobo-wm12-embodied
language: en
---

# Huawei Cloud Open-Sources CloudRobo-WM1.2, a World Model for Robots Trained on Domestic Ascend Compute

> Huawei Cloud released and open-sourced CloudRobo-WM1.2, an Ascend-native embodied world model that topped RoboTwin 2.0 and LiberoPro 2.0 benchmarks — and the open playbook may matter more than the leaderboard rank.

A robot can ace every simulation benchmark thrown at it and still fail on a real factory floor the moment a camera gets partially blocked or the viewing angle shifts. Huawei Cloud (华为云) opened its November 16 launch event in Hangzhou with exactly that problem: in some research tests, model accuracy on a task drops from 62.72% to 7.62% when the viewpoint changes. The company's answer is CloudRobo-WM1.2, what the company calls its first Ascend-native embodied world model — released and open-sourced on the spot.

## What a world model changes for robots

Most robots in production today still run on visual recognition pipelines: see the object, classify it, execute a rehearsed motion. The failure mode is familiar to anyone who has watched a demo robot freeze mid-task. Wu Xiaodong (吴小东), president of Huawei Cloud's embodied product line and director of its Embodied Intelligence Lab, described the pattern bluntly at the event: traditional visual representations go blind when the camera is occluded, confused when the object moves, and lost when the angle changes.

A world model (世界模型) takes a different route. Instead of only recognizing what is in front of the camera, the model predicts how the scene will change as the robot acts — where the cup will swing when gripped, whether the tray edge will collide with the arm, how force transfers through the object. Huawei Cloud calls this moving from "reading a static photo" to "watching a 4D action film in progress."

CloudRobo-WM1.2 packages that idea into five core technologies:

- **PPA multimodal world representation** — perception, prediction, and action unified in one model, so the robot forecasts the physical result of its own motion before committing.

- **DualAct architecture** — the world model itself outputs high-precision actions natively, with a second expert channel for efficient, low-latency control. Developers can size the action expert to their needs.

- **Multi-granularity adaptive pretraining** — the training framework organizes supervision around whatever information each dataset actually contains, so data missing depth, action labels, or camera angles is no longer discarded.

- **Cross-embodily motion modeling** — a shared "motion translator" for single-arm, dual-arm, and wheeled robots, letting one model serve different body types.

- **Four elastic inference modes** — World/Expert Action and Joint/Action-only combinations that teams can match to their compute and latency budgets.

## The numbers behind the launch

The model ranked first on both RoboTwin 2.0 and LiberoPro 2.0, two widely used embodied-AI simulation benchmarks. Huawei Cloud also disclosed some ecosystem figures: it says 80% of Chinese embodied-intelligence (具身智能) companies now build on Ascend (昇腾) cloud, and its CloudRobo platform — launched in June 2025 — has gathered more than 16,500 professional developers and over 230 companies and universities. In June 2026 the platform added an "Industry AI Dream Factory" (行业AI梦工厂) embodied-intelligence zone aimed at shortening the path from model to deployment.

Chief model expert Liu Liu (刘流) walked through the data problem during the event. Collecting effective real-robot data costs roughly 500 to 2,000 yuan per hour — about US$70 to US$280 — so 10,000 hours of usable data can run between about US$700,000 and US$2.8 million. Worse, much of what teams already collect is incomplete: missing depth, missing action annotations, video and motion out of sync. CloudRobo-WM1.2's adaptive pretraining framework is designed to squeeze value from that imperfect data rather than throw it away. The team started experiments with roughly 8,000 hours of pretraining data and has since scaled to about 27,500 hours, drawing from 20 data sources and 33 robot embodiments.

Training ran on domestic Ascend compute rather than Nvidia GPUs — a detail Huawei Cloud leaned on heavily, positioning the release as proof that the full world-model pipeline, from video and 3D geometry to action sequences, can run on China's own silicon.

## Openness with two doors

The open-source announcement has an unusual two-tier structure. Model weights, inference code, and post-training code are public, so individual developers can reproduce and extend the model today. The training framework and the WAM Ascend compute codebase — the engineering practices tuned to Ascend hardware — are opened to core partners first. Huawei Cloud also announced compute grants and industry dataset co-building programs, and paired the release with a concrete product: a next-generation intelligent street-sweeping robot co-developed with sanitation equipment maker Fulongma (福龙马) started pre-sale on launch day, with a full launch event scheduled for December in Shanghai. Ecosystem partners named at the event include listed companies Yijiahe (亿嘉和), Topstar (拓斯达), Fulongma, iSoftStone (软通动力), and Lingyun Guang (凌云光), alongside startups such as Galaxea, Youibot, and Standard Robots.

Wu Xiaodong addressed the obvious tension — why should a Huawei model not compete with the robot makers it supplies? His answer: Huawei Cloud builds the foundation layer, not the endpoint, and partners should be able to "arrive with hardware and leave with a model."

## Honest limitations

First-place rank on two simulation benchmarks is not first place on a production line. Simulation success validates the architecture's theory; long-tail real-world conditions, cloud-to-edge transmission delays, and hardware safety margins still need large-scale physical testing that no leaderboard measures. The benchmark results are also Huawei Cloud's own reported standings, not independently audited. Second, the open release is partial: the Ascend-tuned training framework reaches core partners first, so a team outside that circle can run the model but not fully replicate the training pipeline. Third, the data-efficiency claims rest on the team's own experiments; independent replication on different robot fleets has not yet happened. Finally, the "80% of Chinese embodied-AI companies use Ascend cloud" figure is a vendor disclosure about its own platform, useful as a signal but not a market audit.

## What readers can do now

- **Read the leaderboard context before the headline.** If you are evaluating the model, check what RoboTwin 2.0 and LiberoPro 2.0 actually test — tabletop manipulation and simulation — and plan your own physical validation before any deployment decision.

- **Try the open artifacts.** Weights and inference code are public; a small fine-tuning run on your own robot's footage is the fastest way to see whether the cross-embodily claims hold for your hardware.

- **Watch the December Fulongma launch.** It is the first commercial deployment signal: if the sweeping robot performs in unpredictable outdoor conditions, the world-model thesis gains real evidence beyond simulation.

- **Factor the data economics into your roadmap.** At US$70–280 per hour of real-robot data, the adaptive-pretraining angle matters most for teams sitting on large stores of imperfect legacy data.

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-model-huawei-cloudrobo-wm12-embodied
Free to quote with attribution and a link to the original.
