EgoVerse: Li Feifei's Student Launches International Data Standard for Embodied AI; Guanglun Intelligence is the Only Chinese Company

EgoVerse is a collaborative framework and standard for first-person human manipulation data, initiated by Danfei Xu, a former PhD student of Li Feifei and now an assistant professor at Georgia Tech. The ecosystem currently includes 1362 hours of human demonstration data, about 80,000 episodes, 1965 tasks, 240 scenes, and 2087 collectors. Partners include Georgia Tech, Stanford, ETH Zurich, UC San Diego, Meta, Scale AI, Guanglun Intelligence, Mecka AI, MicroAGI, and Trace Labs. Guanglun Intelligence is the only Chinese company involved. EgoVerse is designed not as a static dataset but as a living framework with common data structures, task semantics, and evaluation methods. Each partner contributes unique expertise: Georgia Tech provides organization and research framework; Stanford's REAL Lab developed the Universal Manipulation Interface (UMI) to transfer human operations to robots; Meta contributes Project Aria data collection glasses; Mecka AI enables data collection via iPhone and lightweight headset; Scale AI provides large-scale data processing; Guanglun Intelligence offers a quality control system across end-side collection, cloud processing, and simulation verification. Guanglun's system includes agent-driven real-time quality control, a unified cloud platform compatible with various hardware, and simulation-based evaluation using its SimFoundry and RoboFinals platforms. This quality loop ensures data is valuable for robot learning. Guanglun is also the only Chinese company participating in both EgoVerse and the Newton physical simulation infrastructure led by NVIDIA, Google DeepMind, and Disney Research.