What is MANSION?

MANSION is designed around a simple idea: long-horizon tasks become much more realistic (and harder) when an agent must operate across multiple floors with persistent object state.

Key goals:

• Long-horizon: tasks that require dozens to hundreds of low-level actions
• Multi-floor: stair/elevator transitions, floor-specific visibility, and navigation costs
• Reproducible: scenes and task definitions are versioned; results are repeatable
• Composable: a skills layer so planners can reason at a higher level than primitive actions

If you’re building LLM/VLM planners, hierarchical RL, or program-of-thought agents, MANSION is meant to be the substrate that makes failures interesting.