interior-env / README.md
Interior Env
Optimize dense reward shaping and update README for Hackathon criteria
2f24049
|
Raw History Blame Contribute Delete
2.67 kB
metadata
title: Spatial Interior Engine
emoji: ๐Ÿ›‹๏ธ
colorFrom: blue
colorTo: purple
sdk: docker
app_port: 7860
tags:
  - openenv
pinned: false
license: mit

Spatial Interior Sandbox (OpenEnv Submission)

A highly advanced 2D Spatial Engine combining "Furniture Tetris" with architectural logic constraints and "A* Pathfinding" for Human Flow validation.

๐ŸŒ Real-World Utility (Domain)

This environment accurately evaluates an AI agent's ability to manipulate physical objects in a 2D coordinate space while preserving accessibility pathways and honoring rigorous interior design constraints. Floor planning, constraint-based packing, and layout flow optimization are genuine robotics and architectural planning problems, moving far past toy games into real-world spatial physics evaluation.

๐Ÿ” Observation Space

The observation space provides a completely typed dictionary of the entire room state:

  • room_size: Tuple (width, height) defining the physical 2D grid.
  • door_position: Fixed entry point coordinate.
  • windows: List of (x, y) coordinates acting as structural constraints.
  • unplaced_items: Inventory of remaining {id, width, height, rules} items.
  • placed_items: Dictionary mapping item IDs to PlacedFurniture(x, y, rotation).
  • feedback_logs: Real-time compiler feedback indicating violations per step.

๐ŸŽฎ Action Space

Agents submit typed JSON payloads to execute positional maneuvers.

{
  "action_type": "place_item", // Or 'remove_item' / 'no_op'
  "item": "sofa",
  "x": 4,
  "y": 5,
  "rotation": 90
}

๐Ÿ† Reward Design & Difficulty

  • Varying Dense Progress Signals: The environment returns a partial reward (+0.5 to +1.0) for every successful, non-overlapping placement that instantly resolves pending architectural rules. Backtracking drops the reward by -0.1, and collisions hit -1.0.
  • 1.0 Perfect Completion Score: Total episode score scales to 1.0 if the room layout is cleanly solved in minimal steps, degrading slowly for wasted actions (efficiency scaling).
  • Difficulty Grading: We scale complexity via easy (8 items), medium (10 items), and hard (15 items) tasks, measuring true spatial capability rather than pure luck. (Random placing fails at 0%).

๐Ÿ› ๏ธ Setup & Inference Evaluator

Fully supported inference.py script provided in the repository root for automated valuation.

  1. Install SDK Tools: run uv lock && pip install openenv-core
  2. Execute Agent Benchmark:
$env:HF_TOKEN="your_hf_token"
python inference.py --task hard
  1. Execute Interactive Dashboard UI: python -m uvicorn server.app:app --port 7860