Newshot

Ascend compute trained a dual-leaderboard SOTA embodied world model, and Huawei Cloud now wants to open up the entire stack

AI summary

The maturing of domestic compute has pulled embodied AI competition away from "how powerful the chip is" and into a multi-dimensional battle over training efficiency, data scale, deployment cost, and how fast scenarios get implemented.

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

Ascend Compute Trained an Embodied World Model That Tops Two Leaderboards, and Huawei Cloud Wants to Open Up the Entire Stack

The maturation of domestic compute has shifted the dimension of embodied AI competition from "how powerful the chip is" into a multi-dimensional battlefield of training efficiency, data scale, deployment cost, and speed of real-world deployment.

Tian Yanlin, reporting from Quantum位 | QbitAI public account

Quantum位 | QbitAI public account

Huawei Cloud is about to open-source an embodied world model, and it's ranked first on two leaderboards right out of the gate!

The key point: this model runs on domestic Ascend compute.

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

In two embodied intelligence benchmarks, RoboTwin 2.0 and LIBERO-Pro 2.0, its average success rate ranks first.

Huawei Cloud plans to open the model weights, inference code, and post-training code on November 30.

Well now — open-sourcing even the post-training code directly, Huawei's move here is indeed aggressive enough.

If another robotics team wants to teach this model a new task, they don't need to build a training pipeline from scratch — grab the ready-made code, hook up their own data, and they can start debugging and validating right away!

And this time, Huawei Cloud didn't stop at "getting the model to run on Ascend chips." Instead, from model architecture, training framework, to core operators and inference deployment, it performed software-hardware co-optimization around the characteristics of embodied models.

From world model pretraining and reinforcement learning to robots actually doing work, Huawei Cloud has connected the entire chain.

This is exactly the core meaning of "Ascend-native."

Domestic compute can train an embodied world model to rank first on two leaderboards

On its debut, Huawei Cloud's embodied world model CloudRobo-WM1.2 delivered two impressive report cards.

On RoboTwin 2.0 and LIBERO-Pro 2.0, it scored 88.98% and 81.73% respectively, taking first place on both.

This model is trained on Ascend compute, and its training framework, key operators, and inference deployment are all optimized around Ascend.

"Ascend-native" refers to exactly this approach.

The simulation leaderboard results are impressive, and real-robot scenarios hold up to the test just as well.

In independent third-party testing, five robot embodiments took the field across 13 manipulation tasks. From turning knobs and continuous dual-arm handovers to putting toys into baskets, the model's average success rate was 23 percentage points higher than the runner-up.

There's an interesting little detail. When the plug slips out unexpectedly after the robot inserts it, CloudRobo-WM1.2 can dynamically adjust its actions based on the new observation frames and retry the insertion.

This means what it remembers isn't just fixed action trajectories — it also possesses the ability to anticipate the outcomes of physical interactions.

Still, after hearing Huawei Cloud's introduction at the recently held embodied world model CloudRobo launch event, I feel its SOTA performance isn't really a surprise.

First, CloudRobo-WM1.2 put considerable effort into "making robots understand the physical world better."

Conventional models can recognize a cup and locate its spatial position, but what they learn is mostly appearance features from a specific viewpoint.

Once physical interaction is involved — for example, if the robot pushes a cup forward, will the cup slide or topple? — ordinary models find it hard to predict such questions accurately; in scenarios with object occlusion, viewpoint changes, and dynamic object motion, the error becomes even more pronounced.

Huawei Cloud adopted the PPA (Perceive-Predict-Act) multimodal world representation architecture, which targets exactly this problem. It jointly trains perception, prediction, and action in the same stage, simultaneously ingesting multiple signal types: RGB video, DINO semantic features, depth, and 3D point clouds.

So the robot not only knows what the cup is and where it is, but can also understand its spatial structure, motion state, and anticipate the physical results of actions.

Complex physical questions like "how will the cup move after being pushed" can be handled with ease.

The more information the model can learn, the more often action annotations, depth, and camera viewpoints in embodied data can't all be gathered together. If one item is missing, can the rest still be used for training?

During pretraining, Huawei Cloud uses multi-granularity supervision masking to exclude missing or invalid parts from loss computation. Whichever part of the information is reliable, the model learns from that part.

It's like doing a comprehensive exam problem: if the conditions for one sub-question are incomplete and unreadable, we don't just give up on the entire problem — we first use the other clear, usable conditions to continue deriving the solution.

Its practical significance isn't about "tolerating dirty data," but about letting large amounts of imperfect heterogeneous data enter training, avoiding discarding an entire sample just because of missing local information.

Real-robot data collection costs as much as 500–1,000 yuan/hour, and the existing valid data is only on the order of millions of hours. Yet leading vendors judge the demand to be on the order of hundreds of millions of hours.

Under this supply-demand gap, data utilization efficiency is itself an engineering competitiveness.

Huawei Cloud also achieved two breakthroughs in "giving robots stronger manipulation capabilities."

First, dual-path collaboration between scene understanding and action generation — the DualAct architecture.

One path is the action expert branch, responsible for directly generating actions; it's faster and suited to tasks with higher response-speed requirements.

The other relies on the world model backbone "world generator," which predicts future states while understanding the scene; it's more precise and suited to fine, complex tasks that require considering physical changes.

Moreover, both paths are trained with the same set of action objectives, and the size of the expert can be adjusted per task, balancing speed and precision.

Second, embodiment-semantics-preserving cross-embodiment motion modeling, solving the challenge of adapting one model to multiple robot forms.

Different robot embodiments differ enormously: single-arm, dual-arm, and wheeled robots have entirely different structures, like different vehicle models with different control logic; but the action intent behind similar tasks is shared.

Huawei Cloud's technology is essentially installing a "translator" on the robot.

The model preserves the native state representations of robots at 96 dimensions and human motion at 168 dimensions, then uses embodiment tokens to distinguish their respective motion interfaces.

Understanding of objects and scenes can be learned together; how to move is autonomously adapted according to each robot's structural characteristics.

According to Huawei Cloud, training data has now expanded to 27,500 hours, incorporating 20 data sources and 33 embodiments.

Additionally, there's a consensus in embodied models: "thinking accurately" and "acting quickly" are quite hard to achieve simultaneously.

Four "packages" are offered for the inference scheme — World/Expert Action, Joint/Action-only — four elastic inference modes that can be combined as needed.

In other words, you can output only actions, or output both actions and predictions; you can also specify that the action expert handles it, or that the world backbone does.

R&D debugging needs precision, actual deployment needs speed — both can be accommodated. R&D teams can flexibly combine them according to compute and latency requirements, like ordering dishes, suiting every budget.

The heaviest tier's computation is about 11 times that of the lightest tier.

Choosing computation by task is CloudRobo-WM1.2's way of saving compute in model design.

But to train long videos, depth, and actions together, and to have the model produce actions promptly on real robots — how does Ascend handle all this work?

"Ascend-native" — where exactly is the native part?

Let's do the math for a 7B-parameter world action model.

Model parameters, gradients, and optimizer states alone would consume about 112GB of VRAM!

Note that the extra overhead from long videos isn't even counted yet.

The compute appetite of embodied world models is unusually large. Video, depth, 3D point clouds, action sequences — the model has to process them all. Compute and VRAM demands far exceed those of ordinary large models.

On real robots, there's another set of calculations. The robot waits for its next command, the model must produce actions promptly, and inference overhead must be kept in check.

This overhead pulls on the entire training system.

How several cards divide the work, how large data is fed in, how the training framework and operators run on the cards, how multi-branch gradients are reduced, selective recomputation of attention, etc. — all must be tuned together around the Ascend hardware.

If one workbench can't spread everything out, let several benches work together.

Huawei Cloud separates data processing from model computation: the former can be handed to Kunpeng, the latter done by Ascend NPUs.

Multiple Ascend cards are interconnected at high speed through super nodes, long video sequences can be split for computation, and model parameters can be stored separately.

Some intermediate results can be recomputed when needed, spending a few extra steps of computation to free up VRAM.

With the tables set up, the next batch of data still needs to be delivered promptly.

Two channels of video at 4Mbps each, recorded for 8,000 hours, amounts to about 28.8TB of compressed video alone.

During training, it must be read repeatedly, decoded, frames extracted, and the footage aligned with the robot trajectory at the same moment.

Huawei Cloud stores video, joint states, end-effector trajectories, chassis, and task labels in a unified way, indexed by one complete manipulation operation.

Once an action segment is found, the related data can be retrieved together. Frequently used data is cached in advance, and the next batch to be used is read in advance. While the cards are still computing the current batch, the next batch is already on its way, avoiding idle compute.

Article title: Ascend compute has trained an embodied world model that achieves SOTA on two leaderboards, and Huawei Cloud now wants to open up the entire foundation stack

Main text: This is a bit like a restaurant kitchen: popular dishes are prepped in advance—before the next order arrives, the ingredients have already been thawed.

Once the data arrives, the training framework still has to distribute the work evenly. Video, point clouds, and actions take up different amounts of GPU memory; if a batch of samples is poorly divided, one card is already packed full while another is still empty.

Huawei Cloud arranges recomputation separately within the framework, combining long sequences and model parallelism to balance the load across cards. When several cards need to exchange results, they can also use the time while data is still in transit to first compute the parts they can handle.

In the reinforcement learning stage, there is likewise a compute-waiting bottleneck.

The model computes actions, the robot in the simulation has to follow them, and then the new frames and results are sent back before training can continue.

Huawei Cloud and the Yu Chao team at Tsinghua University connected the open-source embodied reinforcement learning framework RLinf to Ascend, and one change is quite interesting.

In a set of LIBERO tests, the model outputs 10 steps of actions at a time, and only the 10th step needs to see new frames. So the frames for the first 9 steps do not need to be rendered frame by frame. The physics simulation steps forward one step at a time, rewards are still computed as usual, and rendering overhead is reduced within the action planning cycle.

The Huawei Cloud team connected repeatedly used operations such as normalization and rotary position encoding to fused operators on Ascend NPUs.

Calculations that originally had to be done in several passes can now, in some cases, be combined on the card. Fewer instructions are issued, and intermediate data is moved fewer times.

Huawei Cloud disclosed that after adaptation, the relevant inference efficiency and training throughput improved by about 2 to 3 times.

Breaking it down, in the tests published by RLinf, the number of actions GR00T N1.5 and N1.6 can compute per unit time increased by 195% to 278%, and StarVLA's training invocation throughput improved by 186% to 240%.

How much faster a full round of reinforcement learning can be still depends on simulation, communication, and parameter updates all being counted together.

These tasks were connected to CloudRobo, Huawei Cloud's embodied intelligence development and runtime platform.

When data is hard to collect, you can first synthesize a batch in simulation, then do pretraining, reinforcement learning, and evaluation on Ascend.

When deploying to a real robot, the four inference modes mentioned earlier can be used. During debugging, future frames and actions can be computed together; during on-site work, only the next segment of actions can be computed.

Once the actions are computed, they still have to be delivered to the robot in time.

Huawei Cloud uses R2C (Robot to Cloud) to connect different robots. On-site frames and states can be transmitted upward, and the actions computed by the model can be sent back.

When short-range wireless communication is needed, it can also be combined with SparkLink to reduce transmission latency.

When encountering new operating conditions, the team will bring the collected work data back to CloudRobo, see at which step the model went wrong, then supplement training and evaluation, and deploy the updated model back.

Good—the model can be trained and can also run on real robots. Once in the factory, which process should the robot take on first? Once in the warehouse, which operating condition most needs supplemental data?

The people who deal with equipment every day know these things best.

Bring partners in

But to truly teach a process to a robot, no one can do it all alone.

Model makers entering a factory have to supplement on-site data and learn the craft. Manufacturing enterprises understand production, but they also need someone to help collect data and train models.

Robot body manufacturers, in turn, have to connect the model to their own robots and adjust while running.

Huawei Cloud wants to bring these groups together on CloudRobo to do projects jointly.

For how data enters, how models are trained, how simulation is tested, and how real robots run, the platform provides API, SDK, CLI, and plugin interfaces.

Each party's models, data collection tools, and training frameworks can all be connected and used.

CloudRobo-WM1.2 will also enter the platform, with weights, inference code, and post-training code planned to be released on November 30.

Once things are built, how do you sell them?

In Huawei Cloud's "Industry AI Dream Factory" embodied intelligence zone, built on Ascend Cloud and CloudRobo, partners can publish datasets, models, toolchains, and solutions.

For example, in simulation engines, Huawei Cloud polishes together with partners, and the finished capabilities can be integrated into CloudRobo and also sold on the cloud marketplace.

Wu Xiaodong, President of Huawei Cloud's Embodied Intelligence Product Line and Director of the Embodied Intelligence Lab, also gave a clear statement: under equal conditions, Huawei Cloud gives priority to recommending partners' models and solutions.

Embodied intelligence is entering a critical window for industrialization, and Ascend Cloud has become the de facto compute foundation and ecosystem entry point in the domestic cloud camp.

Since launching in June, the embodied intelligence zone has carried out cooperation with more than 230 enterprises and universities, attracted more than 16,500 developers, and accumulated more than 1,500 industry assets.

At present, out of every 10 embodied intelligence companies in China, perhaps 8 have chosen Huawei Ascend Cloud.

As a member of the first tier in China's embodied intelligence market, Huawei Ascend Cloud has also achieved broad coverage of key domestic universities.

Zhejiang University, Shanghai Jiao Tong University, and others have already entered deep cooperation with Huawei Cloud. They use Ascend compute, simulation platforms, and open-source code to jointly train models, build datasets, and conduct evaluations, and also co-build courses and training bases.

Industrial deployment is also continuing to accelerate: cooperation with Fulongma around unmanned sanitation, jointly defining products and also working together on marketing and launch; jointly releasing a live-line distribution network operation robot with Yijiahe, for large-scale commercial use in high-risk distribution network operations; jointly advancing mine scenario deployment with Sunward...

From industrial manufacturing, urban services, and emergency rescue to lifestyle services, medical care and elderly care, and education and training, CloudRobo's reach has extended to more than 20 industries.

One more thing

The global AI contest is shifting from the performance of a single chip to the capability of an entire system.

In embodied intelligence, electricity must be supplied, the data from robots doing work must be collected back and usable, and the machines must be delivered to customers and able to continue being modified when they enter new scenarios.

China has power infrastructure and the "East Data, West Computing" initiative as a foundation. Domestic compute for training embodied models is gradually maturing, and what comes next is a contest over who trains faster, can use more data, has lower deployment costs, and lands faster.

Being able to run ahead in a single technology is important, but so are the engineering capability and industrial supporting systems that connect every link.

CloudRobo-WM1.2, trained natively on Ascend, took first place on two international benchmarks, which is quite inspiring.

At least for embodied intelligence vendors, on the training side there is finally more than one path besides NVIDIA.

  • Embodied robots enter the real-machine era, and Huawei Cloud CloudRobo begins to team up and push forward 2026-11-20
  • Lurking on Xiaohongshu for 48h: in the AI era, what are top technical talents talking about? 2026-11-18
  • Large models can all control robots now, so how else can Human Data be pushed? 2026-11-13
  • OpenAI's new funding reaches $11 billion, less than half of the original plan 2026-11-07

Related reading

For large models, look at Coding; for embodied intelligence, look at Picking! Yuanli Lingji has already entered the field ahead of others

Fei-Fei Li releases a new world model, and it can run on a single GPU!

Huawei Cloud's Zheng Yelai: the cloud business is least affected by the ban

"For us, what matters more is to steadily improve our software capabilities."

Not a Silicon Valley follower: several young people doing PhDs are betting on an integrated brain for bipedal humanoids

The robot drives a go-kart through continuous turns, fully autonomous in one continuous shot

Huawei Connect 2025: Huawei Cloud uses a new AI engine to reshape intelligent upgrading across thousands of industries

Agentic AI is rewriting the application boundaries of AI

Qianxun Intelligence's Gao Yang: the RobotGPT-1 stage has arrived, and in 4 years it will reach stage 3.5 | MEET 2025

In 10 years, 10% of humans can have exclusive robots

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 ↗