Qiongche Intelligence Unveils Embodied World Model Noe-0

Qiongche Intelligence recently released a technical preview of its embodied world model, Noe-0. Built upon a world model architecture, Noe-0 integrates modeling of real-world environmental dynamics and robot action generation into a single learning process. Upon receiving visual inputs and task semantics, the model simultaneously performs implicit world imagination and action planning within a shared latent representation, directly outputting robot actions. The entire training pipeline for Noe-0 relies exclusively on "ego-centric" (first-person perspective) data collected by Qiongche’s proprietary data acquisition device, RoboPocket, spanning from pre-training through task-specific fine-tuning to establish an end-to-end ego-centric data pipeline. RoboPocket is a portable, wearable device designed to capture visual information and human manipulation trajectories during real-world interactions. To date, Qiongche Intelligence’s data collection network spans over 50 cities across China, amassing more than 100,000 hours of real-world data. Data collectors operate in diverse real-world environments, capturing long-tail samples characterized by varying materials, object shapes, usage states, and manipulation methods based on the specific spatial layouts, objects, and tasks present on-site.