Huawei Cloud Releases Embodied World Model, Responds to Competition with Robotics Companies

LongbridgeAII'm LongbridgeAI, I can summarize articles.

Huawei Cloud has begun to personally engage in training the "brain" of embodied intelligence.

On November 16, at the Huawei Cloud CloudRobo Embodied World Model launch event, Huawei Cloud officially released the Ascend-native embodied world model CloudRobo-WM1.2 and announced its open-source availability.

This model can predict future changes based on an understanding of the physical environment and directly generate robot actions. According to test results disclosed by Huawei Cloud, it achieved first place in embodied intelligence evaluation benchmarks such as RoboTwin 2.0 and LIBERO Pro 2.0.

However, compared to the technical breakthroughs of the model itself, a more pressing issue is the potential competitive relationship between cloud providers, who are now developing their own embodied intelligence models, and the model companies that originally relied on their computing power and platforms.

Wu Xiaodong, President of Huawei Cloud's Embodied Product Line and Director of the Embodied Intelligence Lab, responded at the launch event: "Will we become direct competitors in the future? No."

According to his explanation, Huawei Cloud's continued investment in embodied model research and development is not intended to compete with partners for the model market, but rather to open up model training experience and engineering practices on the Ascend Cloud, helping more enterprises train their own models.

In Huawei Cloud's view, embodied intelligence is moving from laboratories to real industrial scenarios, but there are still many obstacles to large-scale commercialization.

One prominent contradiction is that while models can perform relatively complex operations in the laboratory, they may not work stably when entering real environments such as factories and warehouses.

Robots not only need to identify objects in front of them but also judge the direction of object movement, anticipate changes after operations, and adjust actions in case of accidents. Compared to processing digital information such as text and images, understanding and manipulating the physical world is significantly more difficult.

To address this issue, the CloudRobo-WM1.2 released by Huawei Cloud this time adopts a technical route of joint modeling for perception, prediction, and action.

Simply put, traditional robot models usually require conversion between environmental understanding and action execution, whereas Huawei Cloud attempts to let the world model directly participate in action generation, enabling robots not only to judge what will happen next but also to take action accordingly.

At the same time, considering the response speed requirements for real-world operations, Huawei Cloud has designed a dual-path action architecture.

For complex tasks, the model can simultaneously predict the future environment and actions for more thorough judgment; for simple, high-frequency operations, it can omit part of the prediction process and output actions directly.

Behind this design lies the unavoidable issue of computing costs in embodied intelligence.

The more complete the world model's predictions, the more computing resources it typically consumes, yet robots must respond promptly when executing tasks in real environments.

Huawei Cloud this time provided four composable inference modes, allowing the model to flexibly adjust computational load across different tasks.

According to data announced at the event, the computational requirements of different modes differ by approximately 10 times. Liu Liu, Chief Model Expert at Huawei Cloud's Embodied Intelligence Lab, also revealed that the current model inference latency is about 90 milliseconds, with the next stage aiming to further compress it to within 60 milliseconds.

In addition to inference efficiency, training data is another major challenge facing embodied intelligence models.

Unlike the vast amounts of text and images available on the internet, real operational data required for robot training is not only costly to collect but also difficult to generalize directly across different robots. For example, even when grasping an object, the methods used by single-arm robots, dual-arm robots, and humans can be completely different. Data collected from different devices also suffers from issues such as missing action labels and incomplete sensor information.

The approach chosen by Huawei Cloud this time is to utilize data from various sources as much as possible and learn motion patterns across different embodiments through a unified world model architecture.

Liu Liu disclosed that CloudRobo-WM1.2 previously achieved corresponding test results based on approximately 8,000 hours of pre-training data, and the current scale of training data has expanded to about 27,500 hours, accessing 20 data sources and 33 types of robot embodiments.

In terms of model capabilities, Huawei Cloud stated that it has formed 26 categories of basic operational skills, covering various tasks such as handling, assembly, material transfer, and flexible material manipulation.

However, real factories have far more complex requirements for robot operation stability, continuous running time, maintenance costs, and even safety than experimental environments.

What Huawei Cloud needs to prove is whether, based on Ascend computing power, it can continuously and economically train embodied intelligence models suitable for industrial needs.

And this is precisely why Huawei Cloud is emphasizing an open ecosystem this time.

According to the announced plan, on November 30, Huawei Cloud will open the model weights, inference, and post-training code for CloudRobo-WM1.2, and selectively open pre-training experience and related technical capabilities to certain partners.

Wu Xiaodong also clearly stated that, under equal conditions, Huawei Cloud will prioritize recommending partners' models to create industry solutions for customers.

By providing general-purpose models, computing power, data processing, and development tools, Huawei Cloud can lower the threshold for industry enterprises to develop embodied intelligence systems, with partners completing product development and delivery for specific scenarios.

From the perspective of the platform ecosystem, Huawei Cloud has already accumulated a certain scale of cooperative resources.

Ding Zhibin, Head of Huawei Cloud's Embodied Intelligence Industrial Ecosystem, revealed that since the launch of the Embodied Intelligence Zone in June this year, Huawei Cloud has partnered with over 230 enterprises and universities, attracted more than 16,500 developers, and accumulated over 1,500 digital assets on the platform.

For subsequent ecosystem construction, Huawei Cloud has also introduced more specific support policies.

Among them, for cooperation projects with universities and research teams, each successfully approved project can receive up to RMB 1 million in computing power support; for embodied intelligence special projects targeting core partners, each partner can apply for up to RMB 2 million in support; industry solution projects can also apply for up to RMB 1 million in support.

In the future, Huawei Cloud will also help partners promote and sell models, datasets, and solutions through channels such as the Cloud Store and Industry AI Dream Factory, and leverage Huawei's existing industry customer network to find application scenarios that can truly be implemented.

The sanitation industry is one case showcased at this launch event.

On November 16, Huawei Cloud announced that the new generation of intelligent sweeping robots, jointly developed with sanitation equipment company Fulongma, has started pre-sales and is scheduled for official release in Shanghai on December 17.

Zhang Guifeng, Chairman of Fulongma, mentioned at the event that the company had invested in autonomous driving and intelligent R&D in recent years. Although some products were usable, there was still a significant gap before they could be considered truly effective.

The cooperation between the two parties essentially combines Huawei Cloud's technical capabilities with the scenario accumulation of manufacturing enterprises.

Overall, Huawei Cloud's independent launch of the world model CloudRobo-WM1.2 serves two purposes: on one hand, it verifies the ability of Ascend computing power to train advanced models, providing reusable technical practices for partners; on the other hand, by open-sourcing the model and toolchain, it lowers the threshold for other enterprises' independent R&D, attracting more model vendors into the Ascend Cloud ecosystem.

From the launch of the CloudRobo Embodied Intelligence Platform in 2025 to the current release and announcement of the open-source world model, Huawei Cloud's layout in the field of embodied intelligence has gradually extended from computing power supply to comprehensive applications in models, development tools, and industries.

Risk Warning and Disclaimer

The market carries risks; investment requires caution. This article does not constitute personal investment advice, nor does it consider the specific investment goals, financial status, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article align with their specific circumstances. Responsibility for investments made based on this content rests solely with the investor.

Source: Wallstreetcn The copyright of this article belongs to the original author/organization.

The views expressed herein are solely those of the author and do not reflect the stance of the platform. The content is intended for investment reference purposes only and shall not be considered as investment advice. Please contact us if you have any questions or suggestions regarding the content services provided by the platform.