Huawei Cloud used its Industry AI Dream Factory event on November 16 to release and open-source CloudRobo-WM1.2, an embodied world model built natively for Ascend compute, alongside the training framework and Ascend-optimized practices needed to run it. The company framed the release as its bid to move embodied AI out of lab demos and into industrial deployment on domestically controlled hardware.
The problem the model targets is well known to anyone who has watched robots fail outside staged environments. Conventional vision models depend heavily on raw pixels, so occlusion can collapse detection F1 scores from 68 percent to 28 percent, and out-of-distribution camera angles can send accuracy from 62.72 percent down to 7.62 percent. Real-robot data remains expensive at roughly 500 to 2,000 yuan per hour, and cross-embodiment transfer often degrades as datasets grow, a paradox Huawei attributes to misaligned action semantics between robot bodies.
CloudRobo-WM1.2 answers with five pieces of architecture. A PPA multimodal world representation fuses RGB video, semantic features, depth and point clouds so the model predicts physical outcomes of interaction rather than merely classifying pixels. A world-model-native DualAct architecture provides two output paths, one where the world model directly emits high-precision actions and one where a separate action expert produces low-latency commands. A multi-granularity supervised pretraining framework tolerates missing action labels, partial dimensions, absent depth and incomplete camera views. Embodiment-semantic-faithful motion modeling acts as a shared "action translator" across single-arm, dual-arm and wheeled platforms. Four composable inference modes, World Action, Expert Action, Joint and Action-only, let deployments trade accuracy against latency and compute.
On benchmarks, the company says the model takes the top spot on the RoboTwin 2.0 leaderboard with an 88.98 percent average and ranks first on LIBERO-Pro 2.0 at 81.7 percent, ahead of entries from Tsinghua-NVIDIA teams and open-source VLA baselines. Both figures are company-reported results on academic simulation leaderboards, not independent evaluations, and sim-to-real gaps in industrial settings remain unmeasured.
Beyond the model itself, Huawei Cloud is packaging the surrounding stack: heterogeneous Ascend compute, petabyte-scale multimodal storage, simulation cold start, cloud-native training, cross-embodiment adaptation, continual learning and safety monitoring, offered as six engineering capabilities on the Ascend cloud platform. The open-source release covers the model, the training framework and the WAM Ascend best practices, consistent with the company's pattern of giving away software to anchor hardware demand.
The commercial push is already visible. Huawei Cloud and sanitation equipment maker Forland jointly launched presales for a new intelligent sweeping robot built on CloudRobo tooling, perception and planning models, an agentic runtime and Ascend cloud compute. In September Forland disclosed a RMB 200 million, three-year contract under which Huawei Cloud exclusively develops and deploys its sanitation autonomous driving system for 3-ton and 6-ton vehicles. Huawei has struck similar platform partnerships with Yijiahe, Topsstar and Youai Zhihe for power inspection and industrial handling, deliberately staying out of robot manufacturing itself.
The release lands in a market where Western labs are converging on their own world-model-plus-action stacks, and where compute provenance increasingly matters to buyers. By binding an embodied foundation model to Ascend silicon and giving it away, Huawei is betting that China's robot integrators will standardize on a domestic stack the same way cloud developers standardized on hyperscaler AI platforms. Whether the company-reported leaderboard margins survive contact with real streets and factory floors is the next test.
Comments (0)
Log in to join the discussion
Log InNo comments yet