Noematrix releases Noe-0: an embodiment World Action Model trained entirely on embodiment-free data

机器之心 · wechat · 2026-08-26

Noematrix (穹彻智能) released Noe-0, an embodied-AI pretrained World Action Model trained entirely on embodiment-free data — no teleoperation data at all. Collectors performed tasks directly in real homes and shops across 50+ cities, yielding hundreds of thousands of hours spanning hundreds of thousands of task types.

Noe-0 uses a World Action Model architecture with video prediction as the core learning objective, converting learned world-state dynamics into robot actions. The team found Pixel Prediction gives implicit counterfactual reasoning and significantly improves cross-embodiment transfer. On the data side, Noematrix built the portable RoboPocket capture device (1000+ concurrent collectors at peak) and an AI-native DataEngine with a DataAgent managing distributed collection, QA and labeling — addressing labeling costs that had exceeded collection costs. Noe-0 validates a full loop from pretraining to real-robot execution without teleop data.

Original post →

More from Embodied

Embodied channel →