原生全模态技术战略闭环，智象（HiDream.ai）发布具身世界模型HiDream-O1-Embodied

TL;DR - HiDream.ai launched HiDream-O1-Embodied, a multimodal embodied world model designed to connect language and visual understanding with robotic prediction and execution. It scored 0.692 and ranked first on RoboColiseum’s robustness benchmark, according to the company-provided announcement.

• The model unifies image, video, 3D, and action representations to support an end-to-end understanding–simulation–execution workflow.
• It combines varied language-command understanding, multi-view visual perception, and training under degraded or incomplete conditions to improve robustness.
• HiDream.ai uses a “real-data foundation plus generative augmentation” strategy, expanding motion-capture samples by varying environments, lighting, and objects while preserving physical constraints.
• The release complements HiDream-O1-World: the earlier model targets interactive world understanding and simulation, while HiDream-O1-Embodied focuses on physical-world operation.