Spaces:
Sleeping
Sleeping
File size: 7,937 Bytes
a03a89b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 | """Convert MiniGrid/BabyAI observations into rich natural language text."""
from __future__ import annotations
from typing import Any
import numpy as np
OBJECT_TYPES = {
0: "unseen",
1: "empty",
2: "wall",
3: "floor",
4: "door",
5: "key",
6: "ball",
7: "box",
8: "goal",
9: "lava",
10: "agent",
}
COLORS = {
0: "red",
1: "green",
2: "blue",
3: "purple",
4: "yellow",
5: "grey",
}
DOOR_STATES = {0: "open", 1: "closed", 2: "locked"}
DIRECTION_NAMES = {0: "east", 1: "south", 2: "west", 3: "north"}
_AGENT_ROW = 6
_AGENT_COL = 3
def _format_object_name(obj_type: str, color: str | None, state: str | None = None) -> str:
if obj_type == "door":
prefix = f"{state} {color}".strip() if color else (state or "door")
return f"a {prefix} door".replace(" ", " ").strip()
if color:
return f"a {color} {obj_type}"
return f"a {obj_type}"
def _relative_position_phrase(rel_row: int, rel_col: int) -> str:
parts: list[str] = []
if rel_row < 0:
steps_ahead = abs(rel_row)
parts.append(f"{steps_ahead} step{'s' if steps_ahead != 1 else ''} ahead")
elif rel_row > 0:
steps_behind = rel_row
parts.append(f"{steps_behind} step{'s' if steps_behind != 1 else ''} behind")
if rel_col < 0:
steps_left = abs(rel_col)
parts.append(f"{steps_left} to your left")
elif rel_col > 0:
steps_right = rel_col
parts.append(f"{steps_right} to your right")
if not parts:
return "at your position"
if len(parts) == 1:
return parts[0]
return f"{parts[0]} and {parts[1]}"
def _describe_cell(grid: np.ndarray, row: int, col: int) -> str:
if row < 0 or row >= grid.shape[0] or col < 0 or col >= grid.shape[1]:
return "a wall boundary"
obj_idx = int(grid[row, col, 0])
color_idx = int(grid[row, col, 1])
state_idx = int(grid[row, col, 2])
obj_type = OBJECT_TYPES.get(obj_idx, "unknown")
color = COLORS.get(color_idx)
if obj_type in {"empty", "floor"}:
return "empty space"
if obj_type == "unseen":
return "unseen area"
if obj_type == "wall":
return "a wall"
if obj_type == "door":
return _format_object_name("door", color, DOOR_STATES.get(state_idx, "closed"))
if obj_type == "lava":
return "lava"
return _format_object_name(obj_type, color)
def _scan_objects(grid: np.ndarray) -> list[dict[str, Any]]:
"""Extract notable interactive objects with relative positions."""
objects: list[dict[str, Any]] = []
for row in range(grid.shape[0]):
for col in range(grid.shape[1]):
obj_idx = int(grid[row, col, 0])
color_idx = int(grid[row, col, 1])
state_idx = int(grid[row, col, 2])
obj_type = OBJECT_TYPES.get(obj_idx, "unknown")
if obj_type in {"unseen", "empty", "wall", "floor", "agent"}:
continue
rel_row = row - _AGENT_ROW
rel_col = col - _AGENT_COL
state = DOOR_STATES.get(state_idx) if obj_type == "door" else None
color = COLORS.get(color_idx)
objects.append(
{
"type": obj_type,
"color": color,
"state": state,
"row": row,
"col": col,
"rel_row": rel_row,
"rel_col": rel_col,
"distance": abs(rel_row) + abs(rel_col),
"direction_desc": _relative_position_phrase(rel_row, rel_col),
}
)
objects.sort(key=lambda item: (item["distance"], item["row"], item["col"]))
return objects
def _describe_immediate_surroundings(grid: np.ndarray) -> str:
"""Describe the nearest cells around the agent."""
ahead = _describe_cell(grid, _AGENT_ROW - 1, _AGENT_COL)
left = _describe_cell(grid, _AGENT_ROW, _AGENT_COL - 1)
right = _describe_cell(grid, _AGENT_ROW, _AGENT_COL + 1)
return (
f"Directly ahead: {ahead}.\n"
f"To your left: {left}.\n"
f"To your right: {right}."
)
def _describe_path_ahead(grid: np.ndarray) -> str:
"""Describe what appears in the straight-ahead lane."""
segments: list[str] = []
empty_run = 0
for row in range(_AGENT_ROW - 1, -1, -1):
cell_desc = _describe_cell(grid, row, _AGENT_COL)
if cell_desc == "empty space":
empty_run += 1
continue
if empty_run > 0:
segments.append(
f"empty space for {empty_run} step{'s' if empty_run != 1 else ''}"
)
empty_run = 0
segments.append(cell_desc)
if cell_desc in {"a wall", "a wall boundary", "unseen area"}:
break
if empty_run > 0:
segments.append(
f"empty space for {empty_run} step{'s' if empty_run != 1 else ''}"
)
if not segments:
return "Looking ahead: no clear information."
if len(segments) == 1:
return f"Looking ahead: {segments[0]}."
return f"Looking ahead: {', then '.join(segments)}."
def _describe_notable_objects(objects: list[dict[str, Any]]) -> str:
"""List visible interactive objects with positions."""
if not objects:
return "Notable objects: none visible."
lines = ["Notable objects:"]
for obj in objects:
name = _format_object_name(obj["type"], obj.get("color"), obj.get("state"))
lines.append(f"- {name} ({obj['direction_desc']}).")
return "\n".join(lines)
def _describe_carrying_status(carrying: Any) -> str:
"""Describe what the agent is currently carrying."""
if carrying is None:
return "You are carrying: nothing."
if isinstance(carrying, dict):
obj_type = carrying.get("type")
color = carrying.get("color")
else:
obj_type = getattr(carrying, "type", None)
color = getattr(carrying, "color", None)
if obj_type is None:
return "You are carrying: an object."
if color:
return f"You are carrying: a {color} {obj_type}."
return f"You are carrying: a {obj_type}."
def _render_ascii_grid(grid: np.ndarray) -> str:
"""Render a compact ASCII view for debugging."""
glyphs = {
"unseen": "?",
"empty": ".",
"wall": "#",
"floor": ".",
"door": "D",
"key": "K",
"ball": "B",
"box": "X",
"goal": "G",
"lava": "L",
"agent": "A",
}
rows: list[str] = []
for row in range(grid.shape[0]):
chars: list[str] = []
for col in range(grid.shape[1]):
obj_type = OBJECT_TYPES.get(int(grid[row, col, 0]), "unknown")
chars.append(glyphs.get(obj_type, "!"))
rows.append("".join(chars))
return "\n".join(rows)
def grid_to_text(
obs: dict[str, Any], carrying: Any = None, include_raw_grid: bool = False
) -> str:
"""Convert MiniGrid raw observation dict to a rich language description."""
grid = obs.get("image")
if grid is None:
return "Mission: unknown.\nObservation is missing grid image."
mission = str(obs.get("mission", "")).strip() or "unknown mission"
direction = int(obs.get("direction", 0))
direction_name = DIRECTION_NAMES.get(direction, "unknown")
if not isinstance(grid, np.ndarray):
grid = np.asarray(grid)
objects = _scan_objects(grid)
parts = [
f"Mission: {mission}",
f"You are facing {direction_name}.",
"",
_describe_immediate_surroundings(grid),
_describe_path_ahead(grid),
_describe_notable_objects(objects),
_describe_carrying_status(carrying),
]
if include_raw_grid:
parts.extend(["", "Raw grid (debug):", _render_ascii_grid(grid)])
return "\n".join(part for part in parts if part is not None)
|