Robotics

Huawei Cloud Unveils CloudRobo-WM1.2, an Ascend Trained Robot World Model Slated for Open Source

Huawei Cloud says its new Ascend trained world action model leads two robot manipulation benchmarks and will ship with weights and training code, though no repository has appeared yet.

T
By TechQuire Daily Staff TechQuire Daily Staff
November 16, 2026 / 7 min read

Embodied artificial intelligence rests on a question that text based models never had to answer: what will the physical world do next, and what should a machine do about it? A world model attempts both. Rather than captioning a scene, it learns how objects, surfaces and contacts respond to force, then uses that learned physics to plan the movements a robot should execute. Huawei Cloud has been building toward that capability for more than a year. It launched the CloudRobo embodied intelligence platform in June 2025, and its executives have repeated a narrow division of labour ever since: Huawei supplies the infrastructure, and partners supply the robot bodies.

The competitive backdrop is not neutral. Overseas embodied brain models, including Google Gemini Robotics 2.5 and Nvidia GR00T 2, have iterated while depending heavily on Nvidia's GPU ecosystem, according to CLS. That dependency is the point of comparison Huawei Cloud wants. CloudRobo-WM1.2 was trained on Ascend, Huawei's own silicon, and the company is presenting the release as evidence that a frontier robotics model can be trained without Nvidia hardware.

Huawei Cloud unveiled the model at an online launch on November 16, 2026, and said it would open source the model's weights, its inference code and its training framework. The event ran from 2 p.m. to 3:30 p.m. Beijing time. Wu Xiaodong, president of Huawei Cloud's embodied product line and director of its Embodied AI Lab, led the launch, while chief model expert Liu Liu presented the technology. The company framed the release as a step in moving embodied intelligence from laboratory demonstrations toward large scale deployment, a phrase its launch materials use repeatedly.

There is a gap between the announcement and the artefact. Pandaily reported on November 16 that Huawei Cloud's launch materials had not given a download location, and that no CloudRobo-WM1.2 weights or code repository were publicly listed at the time of writing. That does not make the open source commitment empty, but it does mean the claim currently rests on a slide rather than on a git remote that anyone can clone.

Key Facts

The model pairs two ideas. One is a shared world model backbone that gives a robot a world representation enhanced by predicting what happens next, so the system is rewarded for anticipating consequences rather than for describing pixels. The other is an action generation architecture called DualAct, which Huawei Cloud describes as world model native. In plain terms, the prediction machinery and the movement machinery are not bolted together after the fact; the same learned model that forecasts physical outcomes also produces the action sequence.

Huawei Cloud lists five technical features. PPA is a multimodal world representation architecture. DualAct handles action generation through variable size expert pathways. A pretraining framework applies supervision at several levels of granularity so the model can absorb datasets with incomplete or mismatched labels. Cross embodiment motion modelling tries to keep the meaning of a movement consistent across different robot bodies, so that a grasp learned on one platform transfers rather than collapses. Finally, four inference modes can be combined on demand: World Action or Expert Action, each in Joint or Action only form, letting developers trade planning depth against speed.

The performance claims are specific. Slides shown at the launch put the model's average score at 88.98 percent on RoboTwin 2.0 and 81.7 percent on LIBERO-Pro 2.0, two simulation test suites used to evaluate how well a model plans and executes manipulation tasks. Huawei Cloud says CloudRobo-WM ranks first on both. Those rankings are the company's own claims. Pandaily noted that it had not independently verified which model version produced the leaderboard entries, and The Robotics Media reported on November 16 that the results had not yet been verified by third party labs.

The ecosystem numbers are the other half of the story. CLS reported on November 16 that 80 percent of Chinese embodied intelligence companies have chosen Ascend Cloud, and that Huawei Cloud ranks first in that market in China, with what it describes as near complete coverage of key universities. The CloudRobo platform claims more than 16,500 professional embodied developers and more than 230 companies and universities, spanning industrial, power, sanitation, medical and elder care scenarios. Partners named at the launch include listed firms such as Fulongma, Yijiahe and Topstar, alongside unlisted robot body makers including JAKA, Qianxun, Galaxea, Youibot, Standard Robots and RealMan.

Huawei Cloud also used the event to show a customer deployment rather than a demo. Fulongma Group said a next generation intelligent street sweeping robot developed with Huawei Cloud is now open for presale, with a launch event planned in Shanghai in December. Alongside the model, ifeng reported on November 16 that Huawei Cloud opened its training framework and a set of Ascend best practices for running world action models, aimed at partners, model vendors and customers. The company has also described compute support of up to RMB 1 million, roughly USD 150,000, per approved project, and promised a next model version in the first quarter of 2027.

Analysis

The bigger picture here is that the benchmark table is the marketing, and the compute layer is the product. Huawei Cloud has said plainly that it does not build robot bodies and wants partners to bring hardware and leave with models. A leading open model is the most efficient way to make Ascend the default place where those partners train, fine tune and deploy. Every team that adopts CloudRobo-WM1.2 inherits a toolchain tuned for Ascend, and switching costs rise with every checkpoint, dataset pipeline and inference configuration built on top of it.

The benchmark claims deserve scrutiny rather than dismissal. RoboTwin 2.0 and LIBERO-Pro 2.0 are simulation suites, and simulated manipulation scores are sensitive to training data composition, evaluation protocol and the exact model version submitted. A first place claim on slides is a reproducibility invitation, not a result. The useful test is whether outside labs can reproduce the score with the released weights, and whether the released inference code matches the configuration used for the leaderboard entry. Until a repository exists, the claim cannot be checked in practice.

The open source promise is likewise a schedule question. Weights, inference code and training framework are three different releases with three different levels of usefulness. Weights alone enable evaluation. Inference code enables reproduction. A training framework plus Ascend best practices enables the thing Huawei Cloud actually wants: other organisations building on its stack. The company has committed to all three, and the credibility of the gesture will be settled by what ships, in what order and under what licence, details launch materials did not specify.

There is also a strategic consistency here that is easy to miss. The platform work, the partner list, the compute subsidy and the promised 2027 ecosystem plan with four cooperation paths all point in one direction. Huawei Cloud is not trying to win a robot race. It is trying to become the layer that every robot maker has to pass through, which is a slower and potentially more durable position than shipping a better arm.

Why It Matters

Embodied intelligence is one of the few AI tracks where compute demand spans simulation, synthetic data generation and training at once, which is why it has been a comfortable stronghold for Nvidia. A credible Ascend trained model moves the argument from inference, where domestic chips already had a foothold, to training, which is the harder and more valuable half. That shift matters well beyond one benchmark, because it changes what procurement teams in China can assume about supply risk when they plan robotics programmes.

The developer funnel matters just as much. More than 16,500 developers and more than 230 companies and universities on one platform is a distribution channel before it is a community. If those developers train on Ascend and deploy through CloudRobo, the platform accumulates the data and feedback loops that improve the next model, and the gap to overseas ecosystems narrows for reasons that have little to do with any single architecture.

For buyers, the practical signal is Fulongma's street sweeping robot going to presale with a December launch event in Shanghai. Sanitation is unglamorous, but it is repetitive, commercially funded and already staffed by machines in many cities, which makes it a reasonable first proof point for a model that claims to generalise across robot bodies.

Next Up

The immediate checkpoint is the repository. Until weights, inference code and training framework appear, CloudRobo-WM1.2 remains an announcement with impressive slides, and independent researchers have nothing to test. The second checkpoint is December, when Fulongma holds its launch event in Shanghai and the presale turns into delivered units.

Huawei Cloud has promised a further model version in the first quarter of 2027 and outlined a 2027 ecosystem strategy built around four cooperation paths. Watch whether the open source release arrives before or after that version, and whether third party labs are able to reproduce the RoboTwin 2.0 and LIBERO-Pro 2.0 numbers. Those two facts will decide whether this was a benchmark coup or the opening of a platform.

Tagged

Comments (0)

No comments yet. Be the first to share your thoughts.