Long-horizon human-object-scene interaction (HOSI) generation is a central challenge for embodied AI, animation, and game development. According to a new arXiv paper, existing human-object interaction (HOI) and human-scene interaction (HSI) models each perform well within their respective domains, but the paper argues they fall short when combined into a single long-horizon task spanning both objects and scenes.
The proposed solution, called World2HOSI, is described as a multi-expert harness. Rather than building a single monolithic model, it leverages the strengths of dedicated HOI and HSI experts, implying that each expert handles the subtasks it is already good at and a higher-level orchestration mechanism combines their outputs. The exact architecture and training details are not fully spelled out in the abstract, but the framing suggests a modular approach to a problem that is generally harder than either subproblem alone.
The paper positions World2HOSI as relevant to embodied AI, animation, and game development, where agents or characters must interact with both objects and the surrounding scene over extended time horizons. The abstract does not report quantitative results, so the contribution presented here is primarily the conceptual framework and the explicit focus on the HOSI setting.