Exclusive | Trillion-Dollar Valued Zhipu Open-Sources Embodied AI Model and 1,000-Hour Robot Dataset for the First Time

Specifically, the research team first designed a structured ECoT annotation system, which uses structured fine-grained reasoning annotations. For every frame of each trajectory, it automatically generates six types of atomic supervision information that are universal across all robot embodiments and not tied to hardware: scene description, progress assessment, long-horizon task planning, standardized atomic actions, target object bounding boxes, and discrete action labels.

This annotation scheme is entirely free from robot hardware constraints and can serve as a unified supervision signal to align visual representations across embodiments, fundamentally resolving the problem of representation fragmentation between different robots.

Second, building on the ECoT supervision mechanism, ZR-0 adopts a dual-system architecture (similar to a large brain and a small brain, comprising a System 1/System 2 cognitive framework), decoupling and coordinating the vision-language backbone with the action generation module.

Among them, System 2 uses Alibaba's Qwen3-VL-2B pretrained multimodal large model (VLM) as the core of perception and reasoning, responsible for processing visual observations and task instructions and generating structured ECoT reasoning to capture an understanding of the current scene and task that is independent of embodied form.

System 1, meanwhile, builds a DiT-architecture-based action expert that receives the VLM's representations and maps them via flow matching into embodiment-specific continuous action segments.