Huawei Open-Sources CloudRobo-WM1.2, an Embodied World Model for Robots

Huawei Cloud released and open-sourced CloudRobo-WM1.2 on November 16, an embodied world model trained on its Ascend chips that the company says tops two robotics benchmarks.

Key Takeaways

  • Huawei Cloud released and open-sourced CloudRobo-WM1.2, an embodied world model, on November 16, 2026.
  • The model was trained on Huawei's Ascend AI chips rather than Nvidia hardware and includes five components covering action prediction, generation and cross-robot movement representation.
  • Huawei claims first-place rankings on the RoboTwin 2.0 and LiberoPro 2.0 benchmarks, though these are company-reported results not yet independently verified.
  • Open-sourcing the model, weights and training framework follows a strategy also used by Nvidia's GR00T and other labs to build developer ecosystems around proprietary compute.

At an online launch running from 2 p.m. to 3:30 p.m. Beijing time on November 16, 2026, Huawei Cloud unveiled CloudRobo-WM1.2 and announced it would release the model's weights, training framework and documentation to the public. The presentation was led by Wu Xiaodong, president of Huawei Cloud's embodied products line, and chief model expert Liu Liu, according to the official Huawei Cloud event page.

CloudRobo-WM1.2 is what the industry calls an embodied world model: a system trained to predict how the physical world will respond to an action and to generate the robot actions needed to produce a desired outcome, rather than simply describing a scene in text. Huawei trained the model on its own Ascend AI processors, positioning the release as evidence that Ascend chips can support large-scale robotics model training without relying on Nvidia hardware.

What's Inside the Model

According to a technical breakdown published by CASRAI, Huawei built the model around five components: a multimodal representation for predicting the physical outcomes of actions, an action-generation module called DualAct that uses variable-size expert pathways, an adaptive pretraining method designed to handle incomplete labels across mismatched datasets, a movement representation meant to preserve the meaning of an action across different robot bodies, and four selectable inference modes — World, Expert Action, Joint, and Action-only — that let developers trade off speed for planning depth.

Huawei claims the model ranks first on the RoboTwin 2.0 and LiberoPro 2.0 benchmarks, two test suites used to evaluate how well a model can plan and execute manipulation tasks in simulation. Those results come from Huawei's own reporting and have not yet been independently verified by third-party labs.

Open-Sourcing as a Competitive Strategy

Huawei's decision to release code, weights and Ascend-specific best practices for free mirrors a pattern already set by Nvidia's GR00T line and smaller players like Runway's Praxis-1: give away the model to seed an ecosystem of developers and hardware partners, then sell the compute and cloud infrastructure that model runs on at scale. For Huawei, that infrastructure is Ascend, a chip line built partly to reduce Chinese reliance on Nvidia amid export controls.

The move also puts Huawei in direct competition with a growing field of embodied world models being released by Chinese and American labs alike, including models from Ant Group's Robbyant unit and Google DeepMind's Gemini Robotics line. Benchmarks like RoboTwin and LiberoPro are becoming the proxy battleground for these claims, even though simulation performance does not always translate cleanly to physical robots operating with real sensors and real friction. Huawei has not announced a timeline for when CloudRobo-WM1.2's code repository will go live or which robot hardware partners, if any, have adopted it so far.

Category: Machine Learning

Tags: embodied AI Huawei open source CloudRobo-WM1.2 world model Ascend chips

Related Articles

Frequently Asked Questions

What is CloudRobo-WM1.2?

It is an embodied world model from Huawei Cloud, designed to predict the physical consequences of robot actions and generate the actions needed to complete a task, trained on Huawei's Ascend chips.

Is CloudRobo-WM1.2 open source?

Yes. Huawei announced it is open-sourcing the model's code, weights, training framework and Ascend deployment documentation.

What benchmarks does CloudRobo-WM1.2 claim to lead?

Huawei says the model ranks first on the RoboTwin 2.0 and LiberoPro 2.0 robotics manipulation benchmarks, based on its own reported results.

Why did Huawei train the model on Ascend chips instead of Nvidia hardware?

Huawei has been developing Ascend as a domestic alternative to Nvidia GPUs, partly to reduce reliance on export-controlled chips, and the release demonstrates Ascend's capability for large robotics model training.