Luyang Hu (胡路杨)

We present a lightweight, VLM-guided pipeline that converts kinematic articulated 3D objects into physically realistic, simulation-ready assets for robot learning in under a minute, and parallelizes to thousands of assets at once. Physical properties are refined from multi-view visual inputs through simulator-in-the-loop feedback, and interaction-readiness is assessed along four axes: simulation stability, teleoperated human-object interaction, visual physics realism, and downstream performance under independently trained vision-language-action (VLA) policies. The result is a scalable, low-effort approach to generating interactive assets.

We present a large-scale household navigation dataset of 2,700+ houses with personalized semantics reflecting resident identities, preferences, and relationships. Using LLM-driven top-down generation, we create environments where robots navigate via user-centric queries (e.g., "the owner’s favorite mug") rather than generic spatial commands. This work bridges spatial navigation and context-aware behavior, enabling human-aligned embodied intelligence in personalized home environments.