Huawei Cloud tops robotics benchmarks with CloudRobo-WM1.2
Huawei Cloud unveiled CloudRobo-WM1.2, an Ascend-native embodied world action model featuring predictive world representations and its DualAct dynamic action generation architecture. The model claimed first place on RoboTwin 2.0 and LIBERO-Pro 2.0 with scores of 88.98% and 81.7%, alongside a commitment to open-source future weights.
Huawei's push into embodied AI proves domestic Chinese silicon stacks can compete at the frontier of physical intelligence, but the open-source pledge means little until weights and inference code actually land.
- –DualAct tackles the action gap: Integrating action decoding directly into the predictive world-model backbone reduces compounding rollout errors compared to decoupled planning pipelines.
- –Non-CUDA foundation training: Developing and training a top-tier robotics foundation model end-to-end on Ascend hardware demonstrates real maturity in Huawei's software and compute stack.
- –Setting high-water marks for bimanual manipulation: Scoring 88.98% on RoboTwin 2.0 and 81.7% on LIBERO-Pro 2.0 underscores strong spatial reasoning and sim-to-real execution across complex multi-task environments.
- –Open-weights credibility test: The robotics community has grown weary of open-source announcements that delay model weights; reproducible code and checkpoints will determine whether CloudRobo becomes an ecosystem staple or a PR footnote.
DISCOVERED
12h ago
2026-11-16
PUBLISHED
13h ago
2026-11-16
RELEVANCE
AUTHOR
thePandaily