Newshot

Embodied robots enter the real-machine era as Huawei Cloud's CloudRobo starts building its team

AI summary

A batch of chips can take turns serving multiple companies, and a single adaptation can benefit more teams. No company needs to buy computing power from scratch, nor re-tread the same engineering pitfalls.

Translation(model translation, not the source's own text)

Embodied Robots Enter the Real-Machine Era, Huawei Cloud CloudRobo Begins Teaming Up to Push Forward

A batch of compute cards can serve multiple companies in rotation, and a single adaptation effort can benefit more teams. Compute doesn't have to be bought from scratch by every company, and engineering pitfalls don't have to be re-stepped on by every company.

Tian Yanlin reporting from Aofeisi, Quantum位 | WeChat Official Account QbitAI

Quantum位 | WeChat Official Account QbitAI

A robot gets the job done in the lab—but can it still do it every day once it reaches the factory? Swap in a different robot, and how much of the previously trained capability remains?

Training a smart model isn't enough; the compute, simulation, and real-machine adaptation behind it all have to keep up.

On November 16, Huawei Cloud released the Ascend-native embodied world model CloudRobo-WM1.2.

From simulation training to adaptation across different robots, this model pulls in an entire set of embodied intelligence engineering challenges.

Two days later, Huawei Cloud held the "Industry AI Dream Factory · Embodied Intelligence Executive Forum" in Shanghai, bringing these hidden behind-the-scenes engineering efforts to the forefront.

From computing infrastructure for Agentic computing, to the one-stop embodied intelligence development platform CloudRobo, to industrial practices such as robot adaptive intelligence and live-line work at height, the forum basically covered all the hurdles robots must clear to go from the lab to real-world sites.

A memorandum was also signed on site. Huawei Cloud and Fujian Xiaozhuang Intelligent plan to apply CloudRobo and Ascend compute to university practical training, letting students get hands-on with robot development.

Model release, industrial practice, university training—these look like several different things, but behind them is one consensus: in embodied intelligence R&D, some things don't need to be done from scratch by every company.

Huawei Cloud's approach is to first take on the work of compute, storage, networking, and low-level adaptation, then use CloudRobo to connect data, training, simulation, and real-machine deployment.

As for the actual needs of different industries, it brings in partners who understand the technology and the field to solve them together. The tools and methods that work are kept as much as possible for the next project to continue using.

But to make later engineering less troublesome, the first hurdle must be cleared first.

Before a robotics company has even started training its model, how much compute to buy and how to spend the money is already headache enough.

The Compute Foundation Must First Support the Ambitions of Embodied Intelligence

AI has moved from competing on models to competing on applications, but compute has always been the entry ticket.

Yet even within AI, different projects have genuinely different appetites for compute.

For ordinary Q&A, you ask a question and the large model answers—that round of the task is basically over.

With an Agent, it's not so simple.

First break down the task, look up information, call tools, and after getting results, keep thinking. If any step goes wrong along the way, you may have to start over.

Huawei Ascend chip product management expert Qian Xinyuan gave a set of numbers: Token consumption has grown about 1,400 times in the past two years, and the Token consumption of Agent tasks is 100 to 1,000 times that of ordinary Q&A.

And when it comes to embodied intelligence, compute has to ramp up even further.

A robotic arm reaches out, grasps, and lifts—it looks like just a few seconds. But in simulation training, is the cup actually gripped securely? Will it slip halfway through?

The simulation software must first compute the result before the training program can adjust the model based on feedback. Data such as video, joint positions, and action records must also be fed in batch after batch.

If the simulation results aren't out, training can only wait. If the data doesn't keep up, the cards can only sit idle.

If an embodied intelligence company really had to build its own cluster, just deciding how many cards to buy would be headache enough.

Buy according to ordinary demand, and when training and simulation collide, experiments have to queue up. Buy according to peak demand, and idle equipment still needs routine operations and maintenance.

Not to mention that after the cards are bought, you still have to set up networking, connect storage, and schedule tasks.

Huawei Cloud plans to take over this work.

Hardware procurement, system adaptation, and long-term operations and maintenance are undertaken uniformly by the cloud.

When an embodied team needs to focus on training models, it draws a batch of Ascend compute; when it wants to run more simulation groups, it adjusts resources by task.

Compute follows however the model iterates. There's no need to buy a whole set of equipment in advance for the busiest few days.

But once the cards are provisioned, they still have to actually be kept busy.

Especially for tasks like embodied intelligence, which must handle both long videos and multiple types of data such as actions.

When training on long videos, the compute unit must repeatedly read data from memory and write intermediate results back.

Memory bandwidth governs how fast data moves. It's like an engine with plenty of horsepower but a fuel line that can't supply fuel—the compute can only wait.

Starting from the 950 series onward, Ascend chips have been raising computing power while also speeding up data transmission. According to the plan, the memory bandwidth of the 960DT can reach up to 9.6TB/s!

Once a single card runs smoothly, putting multiple cards together brings new trouble.

Suppose a long video is split across several cards, and all the others still have headroom, but one card runs out of memory first.

Well, everyone else can only stand by and watch.

Huawei Cloud implemented long-sequence parallelism to balance memory loads across different modalities as much as possible, and also prefetches the next batch of training data in advance to reduce waiting.

For embodied intelligence engineers, these optimizations are very practical.

Wait a little less in training, and you can run one more round of experiments. Who doesn't want to see results sooner and pull fewer all-nighters!!!

Some troubles even occur before the model starts training.

When a new model encounters an operator with no ready-made implementation in the framework, engineers have to develop it, port it, and then test performance.

Huawei Cloud plans to have a dedicated team work with embodied partners on Ascend Cloud adaptation to tune such operators.

Pitfalls stepped on by the previous company give later teams a chance to go around them.

For cloud vendors, this is exactly where the business gets interesting.

A batch of cards can serve multiple companies in rotation, and a single adaptation effort can benefit more teams. Compute doesn't have to be bought from scratch by every company, and engineering pitfalls don't have to be re-stepped on by every company.

But no matter how smoothly things run in the training cluster, the robot still hasn't truly started work.

Once the model is trained, the next hurdle is getting it to do the job well in the real world.

From Model to Real Machine, CloudRobo Connects the Engineering Chain

The model clearly knows how to grasp in simulation, but on the real machine, it reaches out and misses.

This is a scenario almost every embodied intelligence company has experienced.

The score in simulation looks great, but on site it suddenly falls apart. Engineers have to go back and check where exactly the problem occurred.

Huawei Cloud CloudRobo aims to solve exactly this kind of trouble.

It connects data processing, model training, simulation verification, and real-machine deployment into a single chain.

After all, when a robot doesn't do a job well, the problem could be at any step. Checking and fixing from scratch every time—who could stand that?

Take data. Collection methods vary widely, and switching devices may mean different data formats.

CloudRobo uses the R2C protocol and SDK to unify robot access, putting data from different sources into one pipeline for processing, supporting PB-level multimodal data management and fast indexing.

Switching collection methods doesn't require tearing down the whole process—just adjust the mapping for the new embodiment.

Currently, the platform has accumulated more than 1 million navigation trajectories, over 50,000 manipulation trajectories, and more than 50 data processing operators.

CloudRobo puts training, inference, post-training, and evaluation into the same development environment, and also brings together more than 30 third-party models, more than 30 evaluation metrics, and more than 20 specialized skills for embodied development.

The Ascend-native embodied world model CloudRobo-WM1.2, planned for official open source on November 30, will also become a model resource on the platform.

Embodied teams don't have to start everything from zero—they can try ready-made models first, then adjust according to their own robots and tasks.

Having a model isn't enough; you also need to know whether it actually works well.

For example, the robot's success rate at grasping cups suddenly drops. Is it a data problem, or is the new model's policy not quite right?

Engineers can reuse the original evaluation environment, adjust the data and the model separately, and see who is actually to blame.

Once the model is tuned, don't rush to put it on the real machine.

CloudRobo supports importing URDF files that describe robot structure, letting your own robotic arm practice in the virtual world first.

The same cup, moved to a different position, rotated to a different angle, grasped again.

Real-machine trial and error must account for equipment wear and safety; in simulation, you can first filter out likely failure cases, then generate data to feed back into training.

But no matter how familiar it gets in the virtual world, it must ultimately pass the real-machine exam.

That's where the CloudRobo runtime management platform comes in.

It connects cloud models with on-site robots. Long-horizon planning and subsequent training can be placed in the cloud; the edge side supports multi-robot scheduling, while real-time control and safety mechanisms stay as close to the body as possible.

When a robot encounters an anomaly, it's handled on site first. Information such as footage and action trajectories before and after the error is also recorded, making it easier for engineers to troubleshoot and optimize the model.

Friends familiar with AI development might think: data management, training, evaluation—aren't these nothing new?

But robots are different. The model finishing its run isn't the end—behind it is a real physical world.

Can actions learned in simulation be reproduced on the real machine? When something goes wrong on site, how do you get into new training faster?

What CloudRobo wants to do is make this chain break down as little as possible.

When the model gets a new version, the original data pipeline and evaluation environment can still be used; when switching robots, the whole development process doesn't need to be rebuilt.

How does it work in practice? The Shanghai Jiao Tong University team shared their experience.

Shanghai Jiao Tong University has launched an embodied intelligence micro-major. If students had to first find robots, set up simulation, and connect data for every group, just fiddling with tools would be enough to keep them busy for a whole course.

Embodied robots enter the real-machine era, and Huawei Cloud's CloudRobo is starting to team up and build momentum

At the Huawei Cloud training camp, 30 students in the embodied intelligence micro-major used CloudRobo to go all the way from simulation to real machines, and also compared and studied models such as Diffusion, ACT, and Pi0.5.

Afterward, some trainees formed teams to participate in the first CloudRobo Embodied Intelligence Competition and won second place.

For universities, students can spend less time setting up environments and get in touch with real robot tasks more quickly.

However, on the industrial site, there will be even more seams. Switching to another factory or another production line may mean different equipment, workflows, and safety requirements.

Huawei Cloud teams up to put more robots on the job

Some problems really do need people who understand the industry to solve together.

Take Yijiahe, for example. It has already put robots onto live high-altitude live-line work sites. Its Z100 has delivered more than 400 units cumulatively and performs more than 80,000 operations each year.

The Z200Pro, developed in cooperation with Huawei Cloud, is planned to enable robots to autonomously judge more on-site environments and plan part of their movement, reducing reliance on manual teleoperation.

According to Yijiahe, the new solution can keep operation time within 30 minutes and improve efficiency by about 30%, while some data still awaits validation on real sites.

Several other ecosystem partners also shared their progress in cooperation with Huawei Cloud on site.

Liu Yufei, executive vice general manager of the National and Local Center, said that the center is relying on Ascend Cloud to jointly build a heterogeneous robot virtual-real fusion training ground, connect real collected data with simulation-synthesized data, and explore dataset openness and unified evaluation, so as to provide training resources for more robots.

Zhou Quan, vice president of Daxiao Robotics, introduced that Daxiao has completed 0-to-1 deployment in scenarios such as retail, security, and hotels, and has cooperated with Ascend Cloud in framework adaptation, operator development, and training tuning. Next, Daxiao plans to distill actions and task workflows into a skill library and standardized delivery solutions, so as to avoid starting from scratch for similar needs.

Yang Guangming, embodied product director of Hangzhou Tiankuan, mentioned that the company is working with Huawei Cloud to build an end-cloud collaborative inspection architecture and optimize multi-robot scheduling with cloud computing power. Its Cloud Elf embodied inspection solution continues to expand deployment in industrial and campus scenarios through ecosystem channels.

The four partners have different entry points.

Some want to solve high-risk operations, some are figuring out how to deliver at scale, and others are specifically filling the gap in training data.

But at the underlying level, data processing, model adaptation, simulation tools, and training tuning also share many commonalities.

What Huawei Cloud's "Industry AI Dream Factory · Embodied Intelligence Zone" aims to do is connect these scattered technical capabilities and accumulate them into something more enterprises can use.

At the bottom is Ascend Cloud, responsible for providing computing power; above it is the CloudRobo platform, which provides full-process development and operation support spanning data management, model development, simulation validation, real-machine deployment, and continuous iteration, lowering the threshold for innovation and application.

On this technical system, universities, developers, model companies, embodied intelligence enterprises, and industry customers can all bring in the capabilities they excel at.

For example, algorithms researched by universities can be handed to enterprises for validation in real tasks; simulation tools polished by tool companies can enter more robot development processes; and operational experience accumulated by industry customers has the chance to be transformed into data needed for model training and evaluation.

Experience proven by one company can then be picked up by the next. This is also one of the important values of ecosystem cooperation.

Of course, it is not enough to simply put what partners have worked hard to build into an asset library.

Through channels such as Huawei Cloud KooGallery, Huawei Cloud pushes partners' products and joint solutions to more customers.

If a simulation tool can be used across multiple projects, the development team will be more motivated to keep maintaining it. Robot companies can also spend less time on repetitive development and leave their energy for their own scenarios.

Beyond technology, tools, and channels, Huawei Cloud has also put up real money to support embodied intelligence technology R&D and the implementation of industry solutions.

The embodied intelligence special program supports up to 2 million yuan per partner, and the AI industry solution special program supports up to 1 million yuan per solution.

At present, five months after the zone launched, it has carried out cooperation with more than 230 enterprises and universities and accumulated more than 1,500 technical assets.

In the battle for embodied intelligence, up to now, as robots move from exhibition booths into factories, what everyone is competing on is not only getting the job done once, but also doing it stably every day and being able to apply the capability in a different scenario.

Behind this, models must be trainable, real machines must run stably, and costs must be calculated clearly on site. No stage can afford to fail.

  • Lurking on Xiaohongshu for 48h: What are top technical talents talking about in the AI era? 2026-11-18
  • Large models can all control robots now, so how far must Human Data still go? 2026-11-13
  • OpenAI's new funding of $11 billion arrives, less than half of the original plan 2026-11-07
  • To remove the AI flavor, humans have started actively writing bad articles... 2026-11-07

Related Reading

Li Auto's smart driving No. 1 starts a business, settling in Beijing Yizhuang

Real-time interaction, real-time video modification, and exploring spatial video, Vidu S2 makes three big moves at once

Hundreds of millions in financing landed! The earliest domestic embodied company to lay out the "human learning" route redoes embodied intelligence from a human perspective

Embodied intelligence is undergoing a silent but profound route shift

An embodied intelligence startup welcomes a Zhejiang University doctoral supervisor, the robot can fly, and VCs are rushing to invest

N robots working collaboratively, zero manual intervention

π0.7 released, VLA bets on the robot's GPT-3 moment

π0.7: A controllable model with emergent capabilities

Galbot turns "robot performances" into "robots on the job," how strong is the end-to-end large model Galactic Brain

Oh wow, this year's Spring Festival Gala stage welcomed a robot that can do work

The translation is machine-generated and individual terms may be off. Nothing is rewritten or cut; the original is authoritative.

Published by 量子位
This site only summarises and aggregates; copyright stays with the original author. Headlines and summaries follow the source — where they differ, the original wins.
Read the original ↗