Scholarlink

We introduce the Virtual Scene Creator (VSC), an end-to-end pipeline that converts a natural-language (NL) prompt into an executable, fully textured virtual-reality (VR) scene. The workflow couples (i) a conversational agent that turns free-form text into a Meta-Language Defined Structure (MLDS), (ii) the Progressive Constraint-Aware Scene Placement (PCASP) algorithm for collision-aware layout, (iii) an automatic validator that corrects spatial errors before assets are generated, and (iv) two user intervention loops for in-editor layout tweaking and selective asset regeneration. We evaluated VSC on 320 automatically generated classroom and business-booth scenes. With the validator enabled, the pipeline achieved a \(93.9\%\) Volumetric Collision-Free Score, a \(96.6\%\) Normalized Collision-Impact Score, and raised the In-Boundary Score by one percentage point over an unvalidated baseline. Omitting floor and wall elements further increased the Collision-Free Score from \(38.0\%\) to \(72.9\%\) . Compared with a traditional LangChain and Pydantic setup, VSC cuts prompt-engineering and post-processing effort by roughly \(80\%\) while keeping the residual error rate within an acceptable margin. Because MLDS artifacts remain human-readable, designers can reposition placeholders in Unity, regenerate only disliked assets, and iterate within minutes, substantially lowering the entry barrier for rapid VR prototyping.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

One Prompt, Many Polygons: Fully Automated VR Scene Building with User Intervention Loops

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Louis Burk,
  • Johannes Armin Koller,
  • Uwe Wienkop

摘要

We introduce the Virtual Scene Creator (VSC), an end-to-end pipeline that converts a natural-language (NL) prompt into an executable, fully textured virtual-reality (VR) scene. The workflow couples (i) a conversational agent that turns free-form text into a Meta-Language Defined Structure (MLDS), (ii) the Progressive Constraint-Aware Scene Placement (PCASP) algorithm for collision-aware layout, (iii) an automatic validator that corrects spatial errors before assets are generated, and (iv) two user intervention loops for in-editor layout tweaking and selective asset regeneration. We evaluated VSC on 320 automatically generated classroom and business-booth scenes. With the validator enabled, the pipeline achieved a \(93.9\%\) Volumetric Collision-Free Score, a \(96.6\%\) Normalized Collision-Impact Score, and raised the In-Boundary Score by one percentage point over an unvalidated baseline. Omitting floor and wall elements further increased the Collision-Free Score from \(38.0\%\) to \(72.9\%\) . Compared with a traditional LangChain and Pydantic setup, VSC cuts prompt-engineering and post-processing effort by roughly \(80\%\) while keeping the residual error rate within an acceptable margin. Because MLDS artifacts remain human-readable, designers can reposition placeholders in Unity, regenerate only disliked assets, and iterate within minutes, substantially lowering the entry barrier for rapid VR prototyping.