Huawei launched CloudRobo-WM1.2, an Ascend-native embodied world model achieving 88.98% on RoboTwin 2.0 and 81.7% on LIBERO-Pro 2.0, featuring DualAct for direct world-model-to-action inference and four inference modes. The open-source release positions it as a competitive robot foundation model with an integrated training framework.
Embodied AI 101 tracks the latest research at the intersection of AI and physical intelligence.
Robotics, manipulation, world models, and the path from digital intelligence to embodied agents — each episode breaks down recent papers and advances in the field.
We examine the methodology, real-world implications, and path to deployment. Covering robot learning, perception, control, and the fundamental challenge of grounding intelligence in the physical world.
New episodes daily.
CloudRobo-WM1.2: Can DualAct Make Robot World Models Practical?.
CloudRobo-WM: A Simple Multi-Modal World Model for Robot Manipulation • CloudRobo team, Huawei Cloud • Technical report and public release materials • 2026.
Cold Open.
Imagine a robot reaching for a mug. One approach generates a possible future: the gripper approaches, closes, and lifts. Another produces the action sequence directly from a representation trained to understand how scenes evolve.
Those approaches raise different engineering questions. Where should actions be generated? And when should the robot spend computation explicitly predicting the future?.
Huawei’s CloudRobo-WM1.2 puts these choices inside one design. Its public framework supports actions generated within the world-model backbone and actions generated by an optional expert network. Huawei calls this arrangement DualAct.
First, a date correction: the launch occurred on November 16, 2026, not November 16, 2025 as stated in the supplied summary. This episode examines the public architecture, Huawei’s technical documentation, and the released benchmark artifacts associated with that announcement.
The interesting story is how predictive learning becomes a configurable control system—and what an Ascend-native implementation adds to that story.
The Idea.
Start with the model’s interface. CloudRobo-WM takes visual observations, robot state, and a natural-language instruction, then generates continuous action sequences. Huawei documents its CloudRobo-WM-7B-Preview asset as a RoboTwin-post-trained variant based on WM1.2. That asset combines a Wan2.2 video backbone with an action-generation branch in a multimodal Mixture-of-Transformers architecture. Huawei reports pretraining on more than ten thousand hours of data.
The use of a video backbone makes temporal prediction part of the representation-learning recipe. Flow matching supplies the generative mechanism: the model learns to transform noisy candidates into structured outputs conditioned on what the robot observes and what it has been asked to do. For control, the useful output is a continuous action chunk.
Huawei calls its representation architecture PPA. The public diagram separates appearance, geometry, and semantics: RGB and pointmaps use VAE encodings, while DINO supplies semantic features. A shared generator predicts future representations and actions, conditioned on language, robot state, and an embodiment identifier.
DualAct then provides two action routes. World Action decodes action tokens within the world generator itself. Expert Action uses an optional action network that reads observed-scene representations from the generator and emits its own sequence. Both routes organize control around the shared world representation.
This is the important meaning of “world-model-native action.” Action generation becomes one of the model’s own outputs. The design connects predictive representation learning to the control interface rather than requiring a separate video-to-action pipeline.
Huawei’s four advertised inference modes are best understood as two switches: which component produces actions, and whether inference also generates future world predictions. Operationally, the combinations are:.
World Action + Joint: the world generator produces actions and future predictions.
World Action + Action-only: the world generator produces actions without future generation.
Expert Action + Joint: the expert produces actions alongside future predictions.
Expert Action + Action-only: the expert produces actions without future generation.
For a practitioner, those switches separate two decisions that are often bundled together. You can choose where action-modeling capacity lives, then choose whether a particular deployment needs explicit future generation. That is a useful way to organize experiments: action-head design and prediction-at-inference become distinct variables.
Huawei also names data-adaptive, multi-granularity pretraining and embodiment-preserving motion modeling as core ingredients. Their stated purpose is to exploit heterogeneous supervision while preserving the meaning of movements across robot bodies. The engineering motivation is straightforward: pooling data becomes more valuable when the model can distinguish different kinds of motion and supervision.
The released RoboTwin post-training pipeline makes the data story concrete. It consumes clean demonstrations in a LeRobot-format export containing RGB video, depth, and camera intrinsics. Data preparation, training resources, and distributed launch procedures are documented together.
Ascend-native also has an identifiable implementation meaning. The public post-training recipe uses PyTorch with , Huawei’s CANN runtime, and DeepSpeed. The accompanying inference adapter includes a CUDA configuration. Training infrastructure and inference deployment are therefore separate, explicit parts of the release.
That separation matters to teams managing mixed compute environments. A useful robot-model release needs an executable route from data preparation through optimization and evaluation. CloudRobo’s Ascend work addresses that route at the training-system level, where framework compatibility and distributed execution become everyday engineering concerns.
The Result.
The headline number is 88.98% average success on RoboTwin 2.0. This figure is corroborated by the benchmark’s public data and the CloudRobo integration’s reported evaluation results. The leaderboard update identifies the new entry with a WM1.2-specific internal identifier.
The protocol is important. RoboTwin’s board specifies clean demonstrations for training, followed by evaluation in both clean and randomized scenes on an Aloha-AgileX dual-arm setup. CloudRobo appears in the co-training track: one policy trained across the benchmark’s fifty tasks. The headline score averages the clean-to-clean and clean-to-random results.
That gives the number a concrete interpretation. It summarizes successful manipulation across a multitask policy under two evaluation conditions. Huawei’s launch slide places CloudRobo at the top of that average-success ordering, and the public benchmark entry supplies the constituent results behind the advertised average.
There is also a useful sensor-level detail. The released adapter uses three RGB cameras and robot joint state. Its evaluation instructions specify zeroed depth and pointmap inputs while RGB remains active. This distinguishes the geometry available during training from the measured sensor inputs required by this evaluation configuration.
The second headline is 81.7% on LIBERO-Pro 2.0, reported in Huawei’s launch slide. That slide explicitly labels the CloudRobo result World Action, connecting the score to the direct world-model action route.
LIBERO-Pro 2.0 adds a complementary test. Its benchmark design includes static distribution shifts and interventions introduced after execution has started. Examples include changes to object conditions, observation failures, action-execution disruptions, and language restatements. These tests ask a policy to continue manipulating under altered conditions, including disturbances encountered during a rollout.
This makes the benchmark choice relevant to Huawei’s architecture. A predictive representation is intended to help control remain grounded as the scene evolves. An intervention benchmark gives that idea an operational target: continue completing the task after the conditions change.
Read together, the two scores tell a coherent story about the evaluation program. RoboTwin supplies broad multitask manipulation coverage on a dual-arm configuration. LIBERO-Pro 2.0 supplies structured robustness tests. Both measure task completion, bringing the discussion back to the quantity roboticists ultimately care about: whether the commanded behavior succeeds.
The Biggest Weakness.
The biggest weakness is incomplete release-to-result traceability for the full WM1.2 claim.
The announcement describes a named model version, four inference modes, and two leading benchmark results. An outside team needs to connect those claims to specific checkpoints, mode settings, and evaluation recipes.
There is already meaningful public implementation material. XPolicyLab includes CloudRobo inference and Ascend post-training sources, checkpoint-download instructions, and detailed RoboTwin results. Saying that no code or weights exist would be incorrect.
The unresolved issue is the completeness of the versioned package. The main CloudRobo research repository still describes its broader code, data, and model release as forthcoming. Public Hugging Face packages associated with the project retain WM-1.0 names, including separate RoboTwin and LIBERO-Pro2 packages. Those names may reflect internal release conventions; they do not, by themselves, establish which advertised WM1.2 capabilities each artifact implements.
In the material I could verify, RoboTwin has the clearest artifact-to-result path. I could not establish an equally complete, version-matched path covering the LIBERO-Pro 2.0 headline and all four inference combinations.
That matters because selecting a mode is central to the product’s technical proposition. A team choosing Expert Action with action-only inference needs the result associated with that configuration. A World Action leaderboard score cannot automatically answer that deployment question.
This is one reproducibility objection, not evidence that the reported scores are false. The remedy is a versioned release that explicitly connects each supported mode to its checkpoint and evaluation configuration, with commands that reproduce both benchmark claims.
Until that connection is complete, the public evidence supports particular CloudRobo implementations more clearly than it supports the entire advertised WM1.2 operating envelope. The distinction matters precisely because the architecture offers choices.
Verdict.
World-action-model researchers should pay attention to the separation between learning predictive representations and generating predictions during execution.
Fast-WAM provides useful context. Its controlled variants retain video co-training while removing explicit future generation at test time. The authors report that this remains competitive, whereas removing video co-training causes a substantially larger performance drop. That result makes prediction during training and prediction during inference separate scientific questions.
CloudRobo’s advertised operating modes put that distinction into a configurable interface. The engineering direction is appealing: preserve a shared representation, then expose choices about the action-producing component and the outputs generated at runtime. It creates a practical vocabulary for discussing where predictive knowledge should live in a robot policy.
Teams already working with Ascend have an especially concrete reason to care. The project’s public post-training material provides an actual NPU software stack, data preparation procedure, and distributed training recipe. Those are the components that turn a model architecture into a repeatable development workflow.
The open-source significance is also practical. Huawei announced WM1.2’s opening, while the publicly inspectable CloudRobo integration already exposes training and inference entry points alongside checkpoint assets. For researchers, that creates opportunities to inspect interfaces, adapt datasets, and evaluate policies through a shared benchmark integration.
The most useful contribution to carry into your own work is the separation of concerns. Representation learning determines what the system can encode about a scene. The action route determines where control capacity sits. Inference mode determines which predictions the deployment actually requests. Hardware support determines whether the training workflow fits the available infrastructure.
Treat those as explicit design decisions, and the discussion becomes more productive than comparing model names.
Takeaway: CloudRobo-WM1.2 organizes world-action modeling around configurable control choices, with a concrete Ascend training implementation behind the model family.
Outro.
Return to the robot reaching for the mug. The valuable question is where predictive learning earns its place in the control system: in the representation, in the generated future, or in the action network using both.
CloudRobo makes those choices unusually visible. That is the idea worth remembering from this launch.
Watch which action route practitioners keep when adapting the model to their own robots—and whether the preferred route changes with the task. The next interesting result is the operating point people actually choose.
This has been Embodied AI 101.