Embodied agents that can perceive, reason, and act across many 3D worlds are a central goal for artificial general intelligence, according to a new arXiv paper. The authors argue that most current research remains confined to specific tasks in a single environment, limiting progress toward general-purpose embodied behavior.
To address that gap, the paper introduces EmbodiedWorld-200K, an embodied-planning dataset, alongside interactive world models. The combination is designed to support "embodied world agents" that can plan and act in varied 3D settings rather than mastering one narrow benchmark.
The source is a single arXiv abstract, so details on dataset construction, model architecture, and evaluation results are not available in the material. The claims here reflect only the authors' stated motivation and proposed approach.