On November 16, Huawei Cloud released CloudRobo-WM1.2, an Ascend-native embodied world model, according to a report by QbitAI. Two days later, the company held an embodied intelligence forum in Shanghai where it presented the surrounding engineering stack: Ascend compute infrastructure, the CloudRobo development platform, and deployments with partners in high-voltage live-line work, inspection, and university training. The model is scheduled to be open-sourced on November 30 and will be offered as a model resource on the platform, the report says.

Why it matters

The report frames the problem as one of repeated engineering rather than model quality alone. Training a capable policy is only part of the work; compute provisioning, simulation, data pipelines, and adaptation to different robot bodies all have to follow. Huawei Cloud's stated approach is to absorb the infrastructure layer, including compute, storage, networking, and low-level adaptation, and connect data, training, simulation, and real-robot deployment through CloudRobo, so that individual teams do not rebuild the same stack.

Technical details

On compute, Huawei Ascend chip product management expert Qian Xinyuan cited figures showing token consumption grew roughly 1,400 times over the past two years, and that Agent tasks consume 100 to 1,000 times the tokens of ordinary question answering. Embodied workloads add further pressure because simulation must return results before training can adjust the model, and video, joint-position, and action data must be fed in batches.

For long-video training, memory bandwidth determines how fast data moves between compute units and memory. According to the report, Ascend chips from the 950 series onward raise both compute and data-transfer speed, with the planned 960DT reaching up to 9.6TB/s of memory bandwidth. Huawei Cloud also implemented long-sequence parallelism to balance memory load across modalities and to prefetch the next batch of training data.

CloudRobo unifies robot access through the R2C protocol and SDK, supporting petabyte-scale multimodal data management and fast indexing. The platform has accumulated more than 1 million navigation trajectories, over 50,000 manipulation trajectories, and more than 50 data-processing operators. Its development environment combines training, inference, post-training, and evaluation, with over 30 third-party models, more than 30 evaluation metrics, and over 20 embodied development skills. It supports importing URDF files so teams can test robot structures in simulation before real-robot trials, and a runtime management platform connects cloud models to on-site robots, keeping real-time control and safety mechanisms near the robot body.

Partner deployments

Yijiahe has deployed its Z100 robot for high-voltage live-line work, with more than 400 units delivered and over 80,000 operations per year. Its Z200Pro, developed with Huawei Cloud, is planned to judge more site conditions and plan some motion autonomously, reducing reliance on teleoperation. Yijiahe says the new approach keeps operation time under 30 minutes with about 30 percent efficiency improvement, though some figures still require validation in real settings.

Shanghai Jiao Tong University's embodied intelligence micro-major used CloudRobo in a training camp with 30 students who worked from simulation through to real robots and compared Diffusion, ACT, and Pi0.5 models; a student team later placed second in the first CloudRobo embodied intelligence competition. Other partners include the National and Local Center, which is building a heterogeneous robot virtual-real training ground on Ascend Cloud, Daxiao Robotics, which has deployed in retail, security, and hotel scenarios, and Hangzhou Tiankuan, which is building a cloud-edge collaborative inspection architecture.

What to watch

Huawei Cloud says its embodied intelligence zone has operated for five months, cooperating with more than 230 companies and universities and accumulating over 1,500 technical assets. It offers up to 2 million yuan per partner for embodied intelligence projects and up to 1 million yuan per solution for AI industry solutions. The open-sourcing of CloudRobo-WM1.2 on November 30 and the validation of partner efficiency claims in real sites are the next concrete checkpoints.

About the author

QbitAI (CN AI)

Thoai Dispatch Curated — curated source featured on Thoai Dispatch.

More by QbitAI (CN AI)