Zhiyuan open-sourced its first embodied dataset for world models and its in-house system, Genie Envisioner-Sim 2.0 (GE-2.0), topped the WorldArena Challenge, taking first place. GE-2.0 incorporates diverse interaction data in training to improve modeling of real-world physics—interactions that are routine for humans but among the hardest and most critical for robots to learn. Zhiyuan says realistic, rich interaction datasets are a core requirement for robots to model and predict physical environ

Zhiyuan open-sourced its first embodied dataset for world models and its in-house system, Genie Envisioner-Sim 2.0 (GE-2.0), topped the WorldArena Challenge, taking first place. GE-2.0 incorporates diverse interaction data in training to improve modeling of real-world physics—interactions that are routine for humans but among the hardest and most critical for robots to learn. Zhiyuan says realistic, rich interaction datasets are a core requirement for robots to model and predict physical environments and advance world-model research.