Huawei Cloud launches Ascend-native embodied world model CloudRobo-WM1.2, open-sourcing Nov 30
智东西 · wechat · 2026-11-16
Huawei Cloud unveiled CloudRobo-WM1.2, its first Ascend-native embodied world model, and will release weights, post-training code, and pre-training best practices on Nov 30. It ranks first on RoboTwin2.0 and LiberoPro2.0 benchmarks, with live demos of industrial material handling, burr cleaning, and unmanned-retail shelf-stocking tasks.
Five core technologies: PPA multimodal world representation that predicts physical interaction outcomes; adaptive pre-training that exploits incomplete heterogeneous data; a DualAct dual-path action architecture (world-model-native action generation plus action experts); cross-embodiment motion modeling unifying 96-dim robot and 168-dim human motion semantics; and four composable inference modes.
Huawei claims its Ascend cloud is "No.1 in China's embodied AI market" — an ecosystem-penetration claim (80% of Chinese embodied AI companies use it), not third-party revenue ranking. A smart sweeper robot co-developed with Fulongma has started pre-sales.
More from Embodied
- MathWorks open-sources a GitHub repo of robotics resources for arms, rovers and drones — lukas_m_ziegler · 2026-11-16
- Open-source CUDA-accelerated 3D LiDAR localization with 3x faster point cloud accumulation — 4310sy · 2026-11-16
- Noumenal Labs open-sources hybrid-aif, a JAX active inference library for robotics — mjdramstead · 2026-11-16
- Wang Xiang: physical-world data is ~6 orders of magnitude larger than language data — 创业邦 · 2026-11-16
- Pony.ai rolls out 7th-gen Toyota Robotaxi in Guangzhou; Changan boosts test efficiency 300% — 创业邦 · 2026-11-16
- Dev open-sources local Apple Health MCP server that feeds workouts to AI agents and pushes plans back to Apple Watch — PrinzPwnage · 2026-11-16