Huawei Cloud open-sources Ascend-native embodied world model CloudRobo-WM1.2, topping two benchmarks
量子位 · wechat · 2026-11-20
Huawei Cloud released CloudRobo-WM1.2, an Ascend-native embodied world model that ranks first on both RoboTwin2.0 (88.98%) and LIBERO-Pro2.0 (81.73%), with model weights, inference code, and post-training code to be open-sourced on Nov 30.
Technical highlights:
- PPA (Perceive-Predict-Act) multimodal representation fusing RGB video, DINO semantic features, depth, and 3D point clouds to predict physical interaction outcomes (e.g., retrying a slipped plug).
- Multi-granularity supervised masking to train on heterogeneous data with missing annotations — critical given real-robot data costs 500–1000 RMB/hour.
- DualAct dual-path architecture (fast action expert + precise world generator) plus four elastic inference modes, with the heaviest 11x the compute of the lightest.
- Cross-embodiment modeling with native 96-dim robot / 168-dim human representations, adapting one model across 33 robot bodies and 27.5k hours of training data.
Full-stack Ascend co-optimization (memory, supernode interconnect, data prefetch, fused operators) yields 2–3x inference/training throughput; integrating the open-source RL framework RLinf boosted GR00T action throughput 195–278% and StarVLA 186–240%. Huawei's embodied AI ecosystem now spans 230+ enterprises/universities and 16.5k+ developers.
Related event: Huawei Cloud Open-Sources CloudRobo-WM1.2, Tops Robot Benchmarks(2 posts)→
More from Embodied
- GXO's 1,000-robot Apptronik deal is a letter of intent, hinging on 4 years of KPIs — YvesMulkers · 2026-11-20
- NVIDIA's SOMA Retargeter adapts human motion to five humanoid robots, from G1 to AgiBot — jonstephens85 · 2026-11-20
- AFFECT from KAIST makes world models dynamics-faithful via action effect matching — John Won · 2026-11-20
- Mind2Act World: a 180-task benchmark exposing robots' mind-action composition failures — Mind2Act-World · 2026-11-20
- ZeroVLN navigates without SFT of the VLM brain using tool-calling and closed-loop RL — ZYT-AI · 2026-11-20
- Robot Paper Club recordings go live with four post-training talks; World Models edition next — DominiqueCAPaul · 2026-11-20