Spaces:
Running
Running
Commit ·
236e665
1
Parent(s): 3e642eb
Initial deploy: OpenEnv RL environment for origami crease patterns
Browse filesFastAPI + openenv-core environment that serves as an RL training
backend for teaching LLMs to generate valid FOLD crease patterns
from natural language prompts. Includes 11 ground-truth patterns
derived from the Optigami simulator and a multi-component reward
function using graph-based structural comparison.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Dockerfile +12 -0
- README.md +34 -6
- app.py +524 -0
- requirements.txt +5 -0
Dockerfile
ADDED
|
@@ -0,0 +1,12 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
FROM python:3.11-slim
|
| 2 |
+
|
| 3 |
+
WORKDIR /app
|
| 4 |
+
|
| 5 |
+
COPY requirements.txt .
|
| 6 |
+
RUN pip install --no-cache-dir -r requirements.txt
|
| 7 |
+
|
| 8 |
+
COPY . .
|
| 9 |
+
|
| 10 |
+
EXPOSE 7860
|
| 11 |
+
|
| 12 |
+
CMD ["uvicorn", "app:app", "--host", "0.0.0.0", "--port", "7860"]
|
README.md
CHANGED
|
@@ -1,11 +1,39 @@
|
|
| 1 |
---
|
| 2 |
-
title:
|
| 3 |
-
emoji:
|
| 4 |
-
colorFrom:
|
| 5 |
-
colorTo:
|
| 6 |
sdk: docker
|
|
|
|
| 7 |
pinned: false
|
| 8 |
-
|
|
|
|
| 9 |
---
|
| 10 |
|
| 11 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
+
title: Optigami
|
| 3 |
+
emoji: 🦢
|
| 4 |
+
colorFrom: indigo
|
| 5 |
+
colorTo: purple
|
| 6 |
sdk: docker
|
| 7 |
+
app_port: 7860
|
| 8 |
pinned: false
|
| 9 |
+
license: mit
|
| 10 |
+
short_description: OpenEnv RL environment for origami crease patterns
|
| 11 |
---
|
| 12 |
|
| 13 |
+
# Optigami - OpenEnv RL Environment for Origami Crease Patterns
|
| 14 |
+
|
| 15 |
+
An RL training environment built with [OpenEnv](https://github.com/meta-pytorch/OpenEnv) that teaches LLMs to generate valid `.fold` crease patterns from natural language prompts.
|
| 16 |
+
|
| 17 |
+
## API Endpoints
|
| 18 |
+
|
| 19 |
+
- `POST /reset` - Get a new origami prompt
|
| 20 |
+
- `POST /step` - Submit a generated FOLD JSON and receive a reward
|
| 21 |
+
|
| 22 |
+
## Reward Structure
|
| 23 |
+
|
| 24 |
+
| Condition | Reward |
|
| 25 |
+
|---|---|
|
| 26 |
+
| Invalid JSON | -2.0 |
|
| 27 |
+
| Missing required FOLD fields | -1.0 |
|
| 28 |
+
| Valid FOLD with structural comparison | -1.0 to 1.0 |
|
| 29 |
+
|
| 30 |
+
## Usage with OpenEnv Client
|
| 31 |
+
|
| 32 |
+
```python
|
| 33 |
+
from openenv.core import EnvClient
|
| 34 |
+
|
| 35 |
+
env = EnvClient("https://openenv-community-optigami-.hf.space")
|
| 36 |
+
obs = env.reset()
|
| 37 |
+
result = env.step('{"vertices_coords": [...], "edges_vertices": [...], ...}')
|
| 38 |
+
print(result.reward)
|
| 39 |
+
```
|
app.py
ADDED
|
@@ -0,0 +1,524 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import json
|
| 2 |
+
from collections import Counter
|
| 3 |
+
from typing import Any, Dict, List, Optional
|
| 4 |
+
|
| 5 |
+
import numpy as np
|
| 6 |
+
import networkx as nx
|
| 7 |
+
from pydantic import Field
|
| 8 |
+
|
| 9 |
+
from openenv.core import (
|
| 10 |
+
Action,
|
| 11 |
+
Environment,
|
| 12 |
+
Observation,
|
| 13 |
+
State,
|
| 14 |
+
create_fastapi_app,
|
| 15 |
+
)
|
| 16 |
+
|
| 17 |
+
# ---------------------------------------------------------------------------
|
| 18 |
+
# Custom Action / Observation models
|
| 19 |
+
# ---------------------------------------------------------------------------
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
class FoldAction(Action):
|
| 23 |
+
"""An action that contains a generated FOLD JSON string."""
|
| 24 |
+
|
| 25 |
+
fold_json: str = Field(description="A JSON string in the FOLD format")
|
| 26 |
+
|
| 27 |
+
|
| 28 |
+
class FoldObservation(Observation):
|
| 29 |
+
"""Observation returned by the origami environment."""
|
| 30 |
+
|
| 31 |
+
prompt: str = Field(default="", description="The origami prompt to fulfill")
|
| 32 |
+
info: Dict[str, Any] = Field(default_factory=dict, description="Extra info")
|
| 33 |
+
|
| 34 |
+
|
| 35 |
+
# ---------------------------------------------------------------------------
|
| 36 |
+
# Dataset: ground-truth origami crease patterns converted from the Optigami
|
| 37 |
+
# TypeScript codebase (lib/patterns.ts). Each entry has a natural-language
|
| 38 |
+
# prompt and a target FOLD-format dict.
|
| 39 |
+
# ---------------------------------------------------------------------------
|
| 40 |
+
|
| 41 |
+
DATASET: List[Dict[str, Any]] = [
|
| 42 |
+
{
|
| 43 |
+
"prompt": "Create a simple valley fold across the center of a square sheet",
|
| 44 |
+
"target": {
|
| 45 |
+
"vertices_coords": [[-1, 1], [1, 1], [-1, 0], [1, 0], [-1, -1], [1, -1]],
|
| 46 |
+
"faces_vertices": [[0, 2, 3], [0, 3, 1], [2, 4, 5], [2, 5, 3]],
|
| 47 |
+
"edges_vertices": [
|
| 48 |
+
[0, 2], [2, 3], [3, 0], [0, 1], [1, 3],
|
| 49 |
+
[2, 4], [4, 5], [5, 3],
|
| 50 |
+
],
|
| 51 |
+
"edges_assignment": ["B", "V", "F", "B", "B", "B", "B", "F"],
|
| 52 |
+
},
|
| 53 |
+
},
|
| 54 |
+
{
|
| 55 |
+
"prompt": "Make an accordion fold with alternating mountain and valley creases",
|
| 56 |
+
"target": {
|
| 57 |
+
"vertices_coords": [
|
| 58 |
+
[-1, 1], [-0.5, 1], [0, 1], [0.5, 1], [1, 1],
|
| 59 |
+
[-1, -1], [-0.5, -1], [0, -1], [0.5, -1], [1, -1],
|
| 60 |
+
],
|
| 61 |
+
"faces_vertices": [
|
| 62 |
+
[0, 1, 6], [0, 6, 5],
|
| 63 |
+
[1, 2, 7], [1, 7, 6],
|
| 64 |
+
[2, 3, 8], [2, 8, 7],
|
| 65 |
+
[3, 4, 9], [3, 9, 8],
|
| 66 |
+
],
|
| 67 |
+
"edges_vertices": [
|
| 68 |
+
[0, 1], [1, 6], [6, 0], [0, 5], [5, 6],
|
| 69 |
+
[1, 2], [2, 7], [7, 1], [6, 7],
|
| 70 |
+
[2, 3], [3, 8], [8, 2], [7, 8],
|
| 71 |
+
[3, 4], [4, 9], [9, 3], [8, 9],
|
| 72 |
+
],
|
| 73 |
+
"edges_assignment": [
|
| 74 |
+
"B", "V", "F", "B", "B",
|
| 75 |
+
"B", "M", "F", "F",
|
| 76 |
+
"B", "V", "F", "F",
|
| 77 |
+
"B", "B", "F", "B",
|
| 78 |
+
],
|
| 79 |
+
},
|
| 80 |
+
},
|
| 81 |
+
{
|
| 82 |
+
"prompt": "Create a waterbomb base with valley folds on diagonals and mountain folds on midlines",
|
| 83 |
+
"target": {
|
| 84 |
+
"vertices_coords": [
|
| 85 |
+
[-1, 1], [0, 1], [1, 1],
|
| 86 |
+
[-1, 0], [0, 0], [1, 0],
|
| 87 |
+
[-1, -1], [0, -1], [1, -1],
|
| 88 |
+
],
|
| 89 |
+
"faces_vertices": [
|
| 90 |
+
[0, 1, 4], [1, 2, 4],
|
| 91 |
+
[2, 5, 4], [5, 8, 4],
|
| 92 |
+
[8, 7, 4], [7, 6, 4],
|
| 93 |
+
[6, 3, 4], [3, 0, 4],
|
| 94 |
+
],
|
| 95 |
+
"edges_vertices": [
|
| 96 |
+
[0, 1], [1, 4], [4, 0],
|
| 97 |
+
[1, 2], [2, 4],
|
| 98 |
+
[2, 5], [5, 4],
|
| 99 |
+
[5, 8], [8, 4],
|
| 100 |
+
[8, 7], [7, 4],
|
| 101 |
+
[7, 6], [6, 4],
|
| 102 |
+
[6, 3], [3, 4],
|
| 103 |
+
[3, 0],
|
| 104 |
+
],
|
| 105 |
+
"edges_assignment": [
|
| 106 |
+
"B", "M", "V",
|
| 107 |
+
"B", "V",
|
| 108 |
+
"B", "M",
|
| 109 |
+
"B", "V",
|
| 110 |
+
"B", "M",
|
| 111 |
+
"B", "V",
|
| 112 |
+
"B", "M",
|
| 113 |
+
"B",
|
| 114 |
+
],
|
| 115 |
+
},
|
| 116 |
+
},
|
| 117 |
+
{
|
| 118 |
+
"prompt": "Make a preliminary fold (square base) with mountain diagonals and valley midlines",
|
| 119 |
+
"target": {
|
| 120 |
+
"vertices_coords": [
|
| 121 |
+
[-1, 1], [0, 1], [1, 1],
|
| 122 |
+
[-1, 0], [0, 0], [1, 0],
|
| 123 |
+
[-1, -1], [0, -1], [1, -1],
|
| 124 |
+
],
|
| 125 |
+
"faces_vertices": [
|
| 126 |
+
[0, 1, 4], [1, 2, 4],
|
| 127 |
+
[2, 5, 4], [5, 8, 4],
|
| 128 |
+
[8, 7, 4], [7, 6, 4],
|
| 129 |
+
[6, 3, 4], [3, 0, 4],
|
| 130 |
+
],
|
| 131 |
+
"edges_vertices": [
|
| 132 |
+
[0, 1], [1, 4], [4, 0],
|
| 133 |
+
[1, 2], [2, 4],
|
| 134 |
+
[2, 5], [5, 4],
|
| 135 |
+
[5, 8], [8, 4],
|
| 136 |
+
[8, 7], [7, 4],
|
| 137 |
+
[7, 6], [6, 4],
|
| 138 |
+
[6, 3], [3, 4],
|
| 139 |
+
[3, 0],
|
| 140 |
+
],
|
| 141 |
+
"edges_assignment": [
|
| 142 |
+
"B", "V", "M",
|
| 143 |
+
"B", "M",
|
| 144 |
+
"B", "V",
|
| 145 |
+
"B", "M",
|
| 146 |
+
"B", "V",
|
| 147 |
+
"B", "M",
|
| 148 |
+
"B", "V",
|
| 149 |
+
"B",
|
| 150 |
+
],
|
| 151 |
+
},
|
| 152 |
+
},
|
| 153 |
+
{
|
| 154 |
+
"prompt": "Create a Miura-ori tessellation fold pattern on a 2x2 grid",
|
| 155 |
+
"target": {
|
| 156 |
+
"vertices_coords": [
|
| 157 |
+
[-1, 1], [0, 1.2], [1, 1],
|
| 158 |
+
[-1, 0], [0, 0.2], [1, 0],
|
| 159 |
+
[-1, -1], [0, -0.8], [1, -1],
|
| 160 |
+
],
|
| 161 |
+
"faces_vertices": [
|
| 162 |
+
[0, 1, 4], [0, 4, 3],
|
| 163 |
+
[1, 2, 5], [1, 5, 4],
|
| 164 |
+
[3, 4, 7], [3, 7, 6],
|
| 165 |
+
[4, 5, 8], [4, 8, 7],
|
| 166 |
+
],
|
| 167 |
+
"edges_vertices": [
|
| 168 |
+
[0, 1], [1, 4], [4, 0], [0, 3],
|
| 169 |
+
[1, 2], [2, 5], [5, 1], [4, 5],
|
| 170 |
+
[3, 4], [4, 7], [7, 3], [3, 6],
|
| 171 |
+
[6, 7], [7, 4], [5, 8], [8, 7],
|
| 172 |
+
],
|
| 173 |
+
"edges_assignment": [
|
| 174 |
+
"B", "M", "F", "B",
|
| 175 |
+
"B", "B", "F", "M",
|
| 176 |
+
"V", "M", "F", "B",
|
| 177 |
+
"B", "F", "B", "B",
|
| 178 |
+
],
|
| 179 |
+
},
|
| 180 |
+
},
|
| 181 |
+
{
|
| 182 |
+
"prompt": "Design a bird base crease pattern with alternating mountain and valley folds radiating from center",
|
| 183 |
+
"target": {
|
| 184 |
+
"vertices_coords": [
|
| 185 |
+
[-1, 1], [0, 1], [1, 1],
|
| 186 |
+
[-1, 0], [0, 0], [1, 0],
|
| 187 |
+
[-1, -1], [0, -1], [1, -1],
|
| 188 |
+
[-0.5, 0.5], [0.5, 0.5], [-0.5, -0.5], [0.5, -0.5],
|
| 189 |
+
],
|
| 190 |
+
"faces_vertices": [
|
| 191 |
+
[0, 9, 3], [0, 1, 9], [1, 4, 9], [3, 9, 4],
|
| 192 |
+
[1, 10, 4], [1, 2, 10], [2, 5, 10], [4, 10, 5],
|
| 193 |
+
[3, 4, 11], [3, 11, 6], [6, 11, 7], [4, 7, 11],
|
| 194 |
+
[4, 5, 12], [5, 8, 12], [8, 7, 12], [4, 12, 7],
|
| 195 |
+
],
|
| 196 |
+
"edges_vertices": [
|
| 197 |
+
[0, 9], [9, 3], [0, 1], [1, 9], [1, 4], [9, 4],
|
| 198 |
+
[1, 10], [1, 2], [2, 10], [2, 5], [5, 10], [10, 4],
|
| 199 |
+
[3, 4], [3, 11], [11, 6], [6, 7], [7, 11], [11, 4],
|
| 200 |
+
[4, 5], [4, 12], [5, 12], [5, 8], [8, 12], [12, 7], [7, 4],
|
| 201 |
+
],
|
| 202 |
+
"edges_assignment": [
|
| 203 |
+
"V", "M", "B", "M", "M", "V",
|
| 204 |
+
"M", "B", "V", "B", "M", "V",
|
| 205 |
+
"M", "M", "V", "B", "M", "V",
|
| 206 |
+
"M", "V", "M", "B", "V", "M", "M",
|
| 207 |
+
],
|
| 208 |
+
},
|
| 209 |
+
},
|
| 210 |
+
{
|
| 211 |
+
"prompt": "Create a hyperbolic paraboloid (hypar) fold on a 4x4 grid with alternating mountain and valley creases",
|
| 212 |
+
"target": {
|
| 213 |
+
"vertices_coords": [
|
| 214 |
+
[-1, 1], [-0.5, 1], [0, 1], [0.5, 1], [1, 1],
|
| 215 |
+
[-1, 0.5], [-0.5, 0.5], [0, 0.5], [0.5, 0.5], [1, 0.5],
|
| 216 |
+
[-1, 0], [-0.5, 0], [0, 0], [0.5, 0], [1, 0],
|
| 217 |
+
[-1, -0.5], [-0.5, -0.5], [0, -0.5], [0.5, -0.5], [1, -0.5],
|
| 218 |
+
[-1, -1], [-0.5, -1], [0, -1], [0.5, -1], [1, -1],
|
| 219 |
+
],
|
| 220 |
+
"faces_vertices": [
|
| 221 |
+
[0, 1, 6], [0, 6, 5], [1, 2, 7], [1, 7, 6],
|
| 222 |
+
[2, 3, 8], [2, 8, 7], [3, 4, 9], [3, 9, 8],
|
| 223 |
+
[5, 6, 11], [5, 11, 10], [6, 7, 12], [6, 12, 11],
|
| 224 |
+
[7, 8, 13], [7, 13, 12], [8, 9, 14], [8, 14, 13],
|
| 225 |
+
[10, 11, 16], [10, 16, 15], [11, 12, 17], [11, 17, 16],
|
| 226 |
+
[12, 13, 18], [12, 18, 17], [13, 14, 19], [13, 19, 18],
|
| 227 |
+
[15, 16, 21], [15, 21, 20], [16, 17, 22], [16, 22, 21],
|
| 228 |
+
[17, 18, 23], [17, 23, 22], [18, 19, 24], [18, 24, 23],
|
| 229 |
+
],
|
| 230 |
+
"edges_vertices": [
|
| 231 |
+
[0, 1], [1, 6], [6, 0], [0, 5], [5, 6],
|
| 232 |
+
[1, 2], [2, 7], [7, 1], [6, 7],
|
| 233 |
+
[2, 3], [3, 8], [8, 2], [7, 8],
|
| 234 |
+
[3, 4], [4, 9], [9, 3], [8, 9],
|
| 235 |
+
[5, 11], [10, 11], [6, 11],
|
| 236 |
+
[6, 12], [11, 12], [7, 12],
|
| 237 |
+
[7, 13], [12, 13], [8, 13],
|
| 238 |
+
[8, 14], [13, 14], [9, 14],
|
| 239 |
+
[5, 10], [10, 16], [10, 15], [15, 16], [11, 16],
|
| 240 |
+
[11, 17], [16, 17], [12, 17],
|
| 241 |
+
[12, 18], [17, 18], [13, 18],
|
| 242 |
+
[13, 19], [18, 19], [14, 19],
|
| 243 |
+
[15, 21], [15, 20], [20, 21], [16, 21],
|
| 244 |
+
[16, 22], [21, 22], [17, 22],
|
| 245 |
+
[17, 23], [22, 23], [18, 23],
|
| 246 |
+
[18, 24], [23, 24], [19, 24],
|
| 247 |
+
],
|
| 248 |
+
"edges_assignment": [
|
| 249 |
+
"B", "M", "F", "B", "V",
|
| 250 |
+
"B", "V", "F", "V",
|
| 251 |
+
"B", "M", "F", "V",
|
| 252 |
+
"B", "B", "F", "V",
|
| 253 |
+
"F", "M", "M",
|
| 254 |
+
"V", "M", "V",
|
| 255 |
+
"M", "M", "M",
|
| 256 |
+
"F", "M", "B",
|
| 257 |
+
"B", "F", "B", "V", "V",
|
| 258 |
+
"M", "V", "V",
|
| 259 |
+
"V", "V", "V",
|
| 260 |
+
"F", "V", "B",
|
| 261 |
+
"F", "B", "B", "V",
|
| 262 |
+
"M", "V", "V",
|
| 263 |
+
"V", "V", "V",
|
| 264 |
+
"F", "B", "B",
|
| 265 |
+
],
|
| 266 |
+
},
|
| 267 |
+
},
|
| 268 |
+
# --- Additional prompt variations for training diversity ---
|
| 269 |
+
{
|
| 270 |
+
"prompt": "Fold a piece of paper in half horizontally",
|
| 271 |
+
"target": {
|
| 272 |
+
"vertices_coords": [[-1, 1], [1, 1], [-1, 0], [1, 0], [-1, -1], [1, -1]],
|
| 273 |
+
"faces_vertices": [[0, 2, 3], [0, 3, 1], [2, 4, 5], [2, 5, 3]],
|
| 274 |
+
"edges_vertices": [
|
| 275 |
+
[0, 2], [2, 3], [3, 0], [0, 1], [1, 3],
|
| 276 |
+
[2, 4], [4, 5], [5, 3],
|
| 277 |
+
],
|
| 278 |
+
"edges_assignment": ["B", "V", "F", "B", "B", "B", "B", "F"],
|
| 279 |
+
},
|
| 280 |
+
},
|
| 281 |
+
{
|
| 282 |
+
"prompt": "Make a zig-zag fold like a paper fan",
|
| 283 |
+
"target": {
|
| 284 |
+
"vertices_coords": [
|
| 285 |
+
[-1, 1], [-0.5, 1], [0, 1], [0.5, 1], [1, 1],
|
| 286 |
+
[-1, -1], [-0.5, -1], [0, -1], [0.5, -1], [1, -1],
|
| 287 |
+
],
|
| 288 |
+
"faces_vertices": [
|
| 289 |
+
[0, 1, 6], [0, 6, 5],
|
| 290 |
+
[1, 2, 7], [1, 7, 6],
|
| 291 |
+
[2, 3, 8], [2, 8, 7],
|
| 292 |
+
[3, 4, 9], [3, 9, 8],
|
| 293 |
+
],
|
| 294 |
+
"edges_vertices": [
|
| 295 |
+
[0, 1], [1, 6], [6, 0], [0, 5], [5, 6],
|
| 296 |
+
[1, 2], [2, 7], [7, 1], [6, 7],
|
| 297 |
+
[2, 3], [3, 8], [8, 2], [7, 8],
|
| 298 |
+
[3, 4], [4, 9], [9, 3], [8, 9],
|
| 299 |
+
],
|
| 300 |
+
"edges_assignment": [
|
| 301 |
+
"B", "V", "F", "B", "B",
|
| 302 |
+
"B", "M", "F", "F",
|
| 303 |
+
"B", "V", "F", "F",
|
| 304 |
+
"B", "B", "F", "B",
|
| 305 |
+
],
|
| 306 |
+
},
|
| 307 |
+
},
|
| 308 |
+
{
|
| 309 |
+
"prompt": "Create the base for a paper crane",
|
| 310 |
+
"target": {
|
| 311 |
+
"vertices_coords": [
|
| 312 |
+
[-1, 1], [0, 1], [1, 1],
|
| 313 |
+
[-1, 0], [0, 0], [1, 0],
|
| 314 |
+
[-1, -1], [0, -1], [1, -1],
|
| 315 |
+
[-0.5, 0.5], [0.5, 0.5], [-0.5, -0.5], [0.5, -0.5],
|
| 316 |
+
],
|
| 317 |
+
"faces_vertices": [
|
| 318 |
+
[0, 9, 3], [0, 1, 9], [1, 4, 9], [3, 9, 4],
|
| 319 |
+
[1, 10, 4], [1, 2, 10], [2, 5, 10], [4, 10, 5],
|
| 320 |
+
[3, 4, 11], [3, 11, 6], [6, 11, 7], [4, 7, 11],
|
| 321 |
+
[4, 5, 12], [5, 8, 12], [8, 7, 12], [4, 12, 7],
|
| 322 |
+
],
|
| 323 |
+
"edges_vertices": [
|
| 324 |
+
[0, 9], [9, 3], [0, 1], [1, 9], [1, 4], [9, 4],
|
| 325 |
+
[1, 10], [1, 2], [2, 10], [2, 5], [5, 10], [10, 4],
|
| 326 |
+
[3, 4], [3, 11], [11, 6], [6, 7], [7, 11], [11, 4],
|
| 327 |
+
[4, 5], [4, 12], [5, 12], [5, 8], [8, 12], [12, 7], [7, 4],
|
| 328 |
+
],
|
| 329 |
+
"edges_assignment": [
|
| 330 |
+
"V", "M", "B", "M", "M", "V",
|
| 331 |
+
"M", "B", "V", "B", "M", "V",
|
| 332 |
+
"M", "M", "V", "B", "M", "V",
|
| 333 |
+
"M", "V", "M", "B", "V", "M", "M",
|
| 334 |
+
],
|
| 335 |
+
},
|
| 336 |
+
},
|
| 337 |
+
{
|
| 338 |
+
"prompt": "Make a simple diagonal fold from corner to corner",
|
| 339 |
+
"target": {
|
| 340 |
+
"vertices_coords": [[0, 0], [1, 0], [1, 1], [0, 1]],
|
| 341 |
+
"faces_vertices": [[0, 1, 2], [0, 2, 3]],
|
| 342 |
+
"edges_vertices": [[0, 1], [1, 2], [2, 3], [3, 0], [0, 2]],
|
| 343 |
+
"edges_assignment": ["B", "B", "B", "B", "V"],
|
| 344 |
+
},
|
| 345 |
+
},
|
| 346 |
+
]
|
| 347 |
+
|
| 348 |
+
|
| 349 |
+
# ---------------------------------------------------------------------------
|
| 350 |
+
# Reward calculation
|
| 351 |
+
# ---------------------------------------------------------------------------
|
| 352 |
+
|
| 353 |
+
|
| 354 |
+
def _build_graph(fold_data: dict) -> nx.Graph:
|
| 355 |
+
"""Convert FOLD data to a networkx graph for structural comparison."""
|
| 356 |
+
G = nx.Graph()
|
| 357 |
+
verts = fold_data.get("vertices_coords", [])
|
| 358 |
+
edges = fold_data.get("edges_vertices", [])
|
| 359 |
+
assignments = fold_data.get("edges_assignment", [])
|
| 360 |
+
|
| 361 |
+
for i, v in enumerate(verts):
|
| 362 |
+
G.add_node(i, x=float(v[0]), y=float(v[1]))
|
| 363 |
+
|
| 364 |
+
for i, e in enumerate(edges):
|
| 365 |
+
a = assignments[i] if i < len(assignments) else "U"
|
| 366 |
+
G.add_edge(int(e[0]), int(e[1]), assignment=a)
|
| 367 |
+
|
| 368 |
+
return G
|
| 369 |
+
|
| 370 |
+
|
| 371 |
+
def _edge_overlap_score(gen_graph: nx.Graph, target_graph: nx.Graph) -> float:
|
| 372 |
+
"""
|
| 373 |
+
Compute a normalised edge-overlap score based on assignment histogram IoU.
|
| 374 |
+
"""
|
| 375 |
+
gen_assign = sorted(
|
| 376 |
+
[d.get("assignment", "U") for _, _, d in gen_graph.edges(data=True)]
|
| 377 |
+
)
|
| 378 |
+
tgt_assign = sorted(
|
| 379 |
+
[d.get("assignment", "U") for _, _, d in target_graph.edges(data=True)]
|
| 380 |
+
)
|
| 381 |
+
|
| 382 |
+
gen_counts = Counter(gen_assign)
|
| 383 |
+
tgt_counts = Counter(tgt_assign)
|
| 384 |
+
|
| 385 |
+
all_keys = set(gen_counts.keys()) | set(tgt_counts.keys())
|
| 386 |
+
intersection = sum(min(gen_counts[k], tgt_counts[k]) for k in all_keys)
|
| 387 |
+
union = sum(max(gen_counts[k], tgt_counts[k]) for k in all_keys)
|
| 388 |
+
|
| 389 |
+
if union == 0:
|
| 390 |
+
return 0.0
|
| 391 |
+
return intersection / union
|
| 392 |
+
|
| 393 |
+
|
| 394 |
+
def calculate_fold_reward(generated_json: str, target_json: dict) -> float:
|
| 395 |
+
"""
|
| 396 |
+
Multi-component reward for comparing a generated FOLD JSON to a target.
|
| 397 |
+
|
| 398 |
+
Components:
|
| 399 |
+
1. JSON validity (-2.0 penalty if invalid)
|
| 400 |
+
2. Required fields (-1.0 penalty if missing)
|
| 401 |
+
3. Vertex count similarity (0 - 0.3)
|
| 402 |
+
4. Edge count similarity (0 - 0.3)
|
| 403 |
+
5. Assignment overlap IoU (0 - 0.4)
|
| 404 |
+
|
| 405 |
+
Total reward clipped to [-2.0, 1.0].
|
| 406 |
+
"""
|
| 407 |
+
try:
|
| 408 |
+
gen_data = json.loads(generated_json)
|
| 409 |
+
except (json.JSONDecodeError, TypeError):
|
| 410 |
+
return -2.0
|
| 411 |
+
|
| 412 |
+
for field in ("vertices_coords", "edges_vertices"):
|
| 413 |
+
if field not in gen_data:
|
| 414 |
+
return -1.0
|
| 415 |
+
|
| 416 |
+
reward = 0.0
|
| 417 |
+
|
| 418 |
+
# Vertex count similarity (max 0.3)
|
| 419 |
+
gen_v = len(gen_data.get("vertices_coords", []))
|
| 420 |
+
tgt_v = len(target_json.get("vertices_coords", []))
|
| 421 |
+
if tgt_v > 0:
|
| 422 |
+
v_ratio = 1.0 - min(abs(gen_v - tgt_v) / tgt_v, 1.0)
|
| 423 |
+
reward += v_ratio * 0.3
|
| 424 |
+
|
| 425 |
+
# Edge count similarity (max 0.3)
|
| 426 |
+
gen_e = len(gen_data.get("edges_vertices", []))
|
| 427 |
+
tgt_e = len(target_json.get("edges_vertices", []))
|
| 428 |
+
if tgt_e > 0:
|
| 429 |
+
e_ratio = 1.0 - min(abs(gen_e - tgt_e) / tgt_e, 1.0)
|
| 430 |
+
reward += e_ratio * 0.3
|
| 431 |
+
|
| 432 |
+
# Assignment overlap (max 0.4)
|
| 433 |
+
try:
|
| 434 |
+
gen_graph = _build_graph(gen_data)
|
| 435 |
+
tgt_graph = _build_graph(target_json)
|
| 436 |
+
overlap = _edge_overlap_score(gen_graph, tgt_graph)
|
| 437 |
+
reward += overlap * 0.4
|
| 438 |
+
except Exception:
|
| 439 |
+
pass
|
| 440 |
+
|
| 441 |
+
return max(-2.0, min(1.0, reward))
|
| 442 |
+
|
| 443 |
+
|
| 444 |
+
# ---------------------------------------------------------------------------
|
| 445 |
+
# OpenEnv Environment
|
| 446 |
+
# ---------------------------------------------------------------------------
|
| 447 |
+
|
| 448 |
+
|
| 449 |
+
class OrigamiEnv(Environment[FoldAction, FoldObservation, State]):
|
| 450 |
+
"""
|
| 451 |
+
Single-turn RL environment for origami crease-pattern generation.
|
| 452 |
+
|
| 453 |
+
reset() -> returns a natural-language prompt as the observation.
|
| 454 |
+
step() -> accepts FOLD JSON, compares to ground truth, returns reward.
|
| 455 |
+
"""
|
| 456 |
+
|
| 457 |
+
def __init__(self) -> None:
|
| 458 |
+
super().__init__()
|
| 459 |
+
self.dataset = DATASET
|
| 460 |
+
self.current_idx = 0
|
| 461 |
+
self._step_count = 0
|
| 462 |
+
self._episode_id: Optional[str] = None
|
| 463 |
+
|
| 464 |
+
def reset(
|
| 465 |
+
self,
|
| 466 |
+
seed: Optional[int] = None,
|
| 467 |
+
episode_id: Optional[str] = None,
|
| 468 |
+
**kwargs: Any,
|
| 469 |
+
) -> FoldObservation:
|
| 470 |
+
if seed is not None:
|
| 471 |
+
np.random.seed(seed)
|
| 472 |
+
self.current_idx = int(np.random.randint(len(self.dataset)))
|
| 473 |
+
self._step_count = 0
|
| 474 |
+
self._episode_id = episode_id
|
| 475 |
+
return FoldObservation(
|
| 476 |
+
prompt=self.dataset[self.current_idx]["prompt"],
|
| 477 |
+
done=False,
|
| 478 |
+
reward=None,
|
| 479 |
+
info={
|
| 480 |
+
"target_vertex_count": len(
|
| 481 |
+
self.dataset[self.current_idx]["target"]["vertices_coords"]
|
| 482 |
+
),
|
| 483 |
+
"target_edge_count": len(
|
| 484 |
+
self.dataset[self.current_idx]["target"]["edges_vertices"]
|
| 485 |
+
),
|
| 486 |
+
},
|
| 487 |
+
)
|
| 488 |
+
|
| 489 |
+
def step(
|
| 490 |
+
self,
|
| 491 |
+
action: FoldAction,
|
| 492 |
+
timeout_s: Optional[float] = None,
|
| 493 |
+
**kwargs: Any,
|
| 494 |
+
) -> FoldObservation:
|
| 495 |
+
target = self.dataset[self.current_idx]["target"]
|
| 496 |
+
reward = calculate_fold_reward(action.fold_json, target)
|
| 497 |
+
self._step_count += 1
|
| 498 |
+
return FoldObservation(
|
| 499 |
+
prompt="Episode finished.",
|
| 500 |
+
done=True,
|
| 501 |
+
reward=reward,
|
| 502 |
+
info={
|
| 503 |
+
"prompt": self.dataset[self.current_idx]["prompt"],
|
| 504 |
+
"target_vertex_count": len(target["vertices_coords"]),
|
| 505 |
+
"target_edge_count": len(target["edges_vertices"]),
|
| 506 |
+
},
|
| 507 |
+
)
|
| 508 |
+
|
| 509 |
+
def state(self) -> State:
|
| 510 |
+
return State(
|
| 511 |
+
episode_id=self._episode_id,
|
| 512 |
+
step_count=self._step_count,
|
| 513 |
+
)
|
| 514 |
+
|
| 515 |
+
|
| 516 |
+
# ---------------------------------------------------------------------------
|
| 517 |
+
# Create the FastAPI application
|
| 518 |
+
# ---------------------------------------------------------------------------
|
| 519 |
+
|
| 520 |
+
app = create_fastapi_app(
|
| 521 |
+
env=OrigamiEnv,
|
| 522 |
+
action_cls=FoldAction,
|
| 523 |
+
observation_cls=FoldObservation,
|
| 524 |
+
)
|
requirements.txt
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
openenv-core>=0.2.1
|
| 2 |
+
fastapi
|
| 3 |
+
uvicorn
|
| 4 |
+
numpy
|
| 5 |
+
networkx
|