Spaces:
Running
Running
Andy Lee
commited on
Commit
Β·
04ae29a
1
Parent(s):
adf0c49
feat: force model to react with lat and lon for guessing
Browse files- geo_bot.py +118 -58
geo_bot.py
CHANGED
|
@@ -3,6 +3,7 @@ import json
|
|
| 3 |
import re
|
| 4 |
from io import BytesIO
|
| 5 |
from typing import Tuple, List, Optional, Dict, Any, Type
|
|
|
|
| 6 |
|
| 7 |
from PIL import Image
|
| 8 |
from langchain_core.messages import HumanMessage, BaseMessage
|
|
@@ -37,7 +38,10 @@ AGENT_PROMPT_TEMPLATE = """
|
|
| 37 |
|
| 38 |
4. **Be Decisive:** A unique, definitive clue (full address, rare town name, etc.) β `GUESS` immediately.
|
| 39 |
|
| 40 |
-
5. **Final-Step Rule
|
|
|
|
|
|
|
|
|
|
| 41 |
|
| 42 |
ββββββββββββββββββββββββββββββββ
|
| 43 |
**Context & Task:**
|
|
@@ -136,21 +140,33 @@ class GeoBot:
|
|
| 136 |
)
|
| 137 |
]
|
| 138 |
|
| 139 |
-
def _parse_agent_response(
|
|
|
|
|
|
|
| 140 |
"""
|
| 141 |
-
Robustly parses JSON from the LLM response
|
| 142 |
"""
|
| 143 |
try:
|
| 144 |
assert isinstance(response.content, str), "Response content is not a string"
|
| 145 |
content = response.content.strip()
|
|
|
|
|
|
|
|
|
|
| 146 |
match = re.search(r"```json\s*(\{.*?\})\s*```", content, re.DOTALL)
|
| 147 |
if match:
|
| 148 |
json_str = match.group(1)
|
|
|
|
| 149 |
else:
|
| 150 |
json_str = content
|
| 151 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 152 |
except (json.JSONDecodeError, AttributeError) as e:
|
| 153 |
-
print(f"
|
|
|
|
| 154 |
return None
|
| 155 |
|
| 156 |
def init_history(self) -> List[Dict[str, Any]]:
|
|
@@ -222,7 +238,8 @@ class GeoBot:
|
|
| 222 |
prompt, image_b64_for_prompt[-1:]
|
| 223 |
)
|
| 224 |
response = self.model.invoke(message)
|
| 225 |
-
|
|
|
|
| 226 |
except Exception as e:
|
| 227 |
print(f"Error during model invocation: {e}")
|
| 228 |
decision = None
|
|
@@ -259,15 +276,7 @@ class GeoBot:
|
|
| 259 |
self, max_steps: int = 10, step_callback=None
|
| 260 |
) -> Optional[Tuple[float, float]]:
|
| 261 |
"""
|
| 262 |
-
|
| 263 |
-
|
| 264 |
-
Args:
|
| 265 |
-
max_steps: Maximum number of steps to take
|
| 266 |
-
step_callback: Function called after each step with step info
|
| 267 |
-
Signature: callback(step_info: dict) -> None
|
| 268 |
-
|
| 269 |
-
Returns:
|
| 270 |
-
Final guess coordinates (lat, lon) or None if no guess made
|
| 271 |
"""
|
| 272 |
history = self.init_history()
|
| 273 |
|
|
@@ -275,14 +284,24 @@ class GeoBot:
|
|
| 275 |
step_num = max_steps - step + 1
|
| 276 |
print(f"\n--- Step {step_num}/{max_steps} ---")
|
| 277 |
|
| 278 |
-
#
|
| 279 |
-
|
| 280 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 281 |
|
| 282 |
-
screenshot_bytes = self.controller.take_street_view_screenshot()
|
| 283 |
if not screenshot_bytes:
|
| 284 |
-
print("Failed to
|
| 285 |
-
return
|
| 286 |
|
| 287 |
current_screenshot_b64 = self.pil_to_base64(
|
| 288 |
image=Image.open(BytesIO(screenshot_bytes))
|
|
@@ -290,36 +309,28 @@ class GeoBot:
|
|
| 290 |
available_actions = self.controller.get_available_actions()
|
| 291 |
print(f"Available actions: {available_actions}")
|
| 292 |
|
| 293 |
-
#
|
| 294 |
-
if step == 1: # Final step
|
| 295 |
-
|
| 296 |
-
|
| 297 |
-
|
| 298 |
-
"action_details": {"action": "GUESS", "lat": 0.0, "lon": 0.0},
|
| 299 |
-
}
|
| 300 |
-
# Try to get a real guess from AI
|
| 301 |
-
try:
|
| 302 |
-
ai_decision = self.execute_agent_step(
|
| 303 |
-
history, step, current_screenshot_b64, available_actions
|
| 304 |
-
)
|
| 305 |
-
if (
|
| 306 |
-
ai_decision
|
| 307 |
-
and ai_decision.get("action_details", {}).get("action")
|
| 308 |
-
== "GUESS"
|
| 309 |
-
):
|
| 310 |
-
decision = ai_decision
|
| 311 |
-
except Exception as e:
|
| 312 |
-
print(
|
| 313 |
-
f"\nERROR: An exception occurred during the final GUESS attempt: {e}. Using fallback (0,0).\n"
|
| 314 |
-
)
|
| 315 |
else:
|
| 316 |
-
# Normal step execution
|
| 317 |
decision = self.execute_agent_step(
|
| 318 |
history, step, current_screenshot_b64, available_actions
|
| 319 |
)
|
| 320 |
|
| 321 |
-
|
| 322 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 323 |
step_info = {
|
| 324 |
"step_num": step_num,
|
| 325 |
"max_steps": max_steps,
|
|
@@ -330,7 +341,7 @@ class GeoBot:
|
|
| 330 |
"is_final_step": step == 1,
|
| 331 |
"reasoning": decision.get("reasoning", "N/A"),
|
| 332 |
"action_details": decision.get("action_details", {"action": "N/A"}),
|
| 333 |
-
"history": history.copy(),
|
| 334 |
}
|
| 335 |
|
| 336 |
action_details = decision.get("action_details", {})
|
|
@@ -338,29 +349,78 @@ class GeoBot:
|
|
| 338 |
print(f"AI Reasoning: {decision.get('reasoning', 'N/A')}")
|
| 339 |
print(f"AI Action: {action}")
|
| 340 |
|
| 341 |
-
# Call UI callback before executing action
|
| 342 |
if step_callback:
|
| 343 |
try:
|
| 344 |
step_callback(step_info)
|
| 345 |
except Exception as e:
|
| 346 |
-
print(f"
|
| 347 |
|
| 348 |
-
# Add
|
| 349 |
self.add_step_to_history(history, current_screenshot_b64, decision)
|
| 350 |
|
| 351 |
# Execute action
|
| 352 |
if action == "GUESS":
|
| 353 |
-
lat
|
| 354 |
-
|
| 355 |
-
|
| 356 |
-
|
| 357 |
-
|
| 358 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 359 |
else:
|
| 360 |
self.execute_action(action)
|
| 361 |
|
| 362 |
-
print("Max steps reached
|
| 363 |
-
return
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 364 |
|
| 365 |
def analyze_image(self, image: Image.Image) -> Optional[Tuple[float, float]]:
|
| 366 |
image_b64 = self.pil_to_base64(image)
|
|
|
|
| 3 |
import re
|
| 4 |
from io import BytesIO
|
| 5 |
from typing import Tuple, List, Optional, Dict, Any, Type
|
| 6 |
+
import time
|
| 7 |
|
| 8 |
from PIL import Image
|
| 9 |
from langchain_core.messages import HumanMessage, BaseMessage
|
|
|
|
| 38 |
|
| 39 |
4. **Be Decisive:** A unique, definitive clue (full address, rare town name, etc.) β `GUESS` immediately.
|
| 40 |
|
| 41 |
+
5. **Final-Step Rule**
|
| 42 |
+
- If **Remaining Steps = 1**, you **MUST** `GUESS` with coordinates.
|
| 43 |
+
- **NO EXCEPTIONS**: Even with limited clues, provide your best estimate.
|
| 44 |
+
- **ALWAYS provide lat/lon numbers** - educated guesses are mandatory.
|
| 45 |
|
| 46 |
ββββββββββββββββββββββββββββββββ
|
| 47 |
**Context & Task:**
|
|
|
|
| 140 |
)
|
| 141 |
]
|
| 142 |
|
| 143 |
+
def _parse_agent_response(
|
| 144 |
+
self, response: BaseMessage, verbose: bool = False
|
| 145 |
+
) -> Optional[Dict[str, Any]]:
|
| 146 |
"""
|
| 147 |
+
Robustly parses JSON from the LLM response with detailed logging.
|
| 148 |
"""
|
| 149 |
try:
|
| 150 |
assert isinstance(response.content, str), "Response content is not a string"
|
| 151 |
content = response.content.strip()
|
| 152 |
+
if verbose:
|
| 153 |
+
print(f"Raw AI response: {content[:200]}...") # Show first 200 chars
|
| 154 |
+
|
| 155 |
match = re.search(r"```json\s*(\{.*?\})\s*```", content, re.DOTALL)
|
| 156 |
if match:
|
| 157 |
json_str = match.group(1)
|
| 158 |
+
print(f"Extracted JSON: {json_str}")
|
| 159 |
else:
|
| 160 |
json_str = content
|
| 161 |
+
print("No JSON code block found, trying to parse entire content")
|
| 162 |
+
|
| 163 |
+
parsed = json.loads(json_str)
|
| 164 |
+
print(f"Successfully parsed JSON: {parsed}")
|
| 165 |
+
return parsed
|
| 166 |
+
|
| 167 |
except (json.JSONDecodeError, AttributeError) as e:
|
| 168 |
+
print(f"β JSON parsing failed: {e}")
|
| 169 |
+
print(f"Full response was:\n{response.content}")
|
| 170 |
return None
|
| 171 |
|
| 172 |
def init_history(self) -> List[Dict[str, Any]]:
|
|
|
|
| 238 |
prompt, image_b64_for_prompt[-1:]
|
| 239 |
)
|
| 240 |
response = self.model.invoke(message)
|
| 241 |
+
verbose = remaining_steps == 1
|
| 242 |
+
decision = self._parse_agent_response(response, verbose)
|
| 243 |
except Exception as e:
|
| 244 |
print(f"Error during model invocation: {e}")
|
| 245 |
decision = None
|
|
|
|
| 276 |
self, max_steps: int = 10, step_callback=None
|
| 277 |
) -> Optional[Tuple[float, float]]:
|
| 278 |
"""
|
| 279 |
+
Agent loop with simple retry logic and clear error coordinates.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 280 |
"""
|
| 281 |
history = self.init_history()
|
| 282 |
|
|
|
|
| 284 |
step_num = max_steps - step + 1
|
| 285 |
print(f"\n--- Step {step_num}/{max_steps} ---")
|
| 286 |
|
| 287 |
+
# Simple retry for screenshot
|
| 288 |
+
screenshot_bytes = None
|
| 289 |
+
for retry in range(3):
|
| 290 |
+
try:
|
| 291 |
+
self.controller.setup_clean_environment()
|
| 292 |
+
self.controller.label_arrows_on_screen()
|
| 293 |
+
screenshot_bytes = self.controller.take_street_view_screenshot()
|
| 294 |
+
if screenshot_bytes:
|
| 295 |
+
break
|
| 296 |
+
print(f"Screenshot retry {retry + 1}/3")
|
| 297 |
+
except Exception as e:
|
| 298 |
+
print(f"Error in step {step_num}, retry {retry + 1}: {e}")
|
| 299 |
+
if retry < 2:
|
| 300 |
+
time.sleep(2)
|
| 301 |
|
|
|
|
| 302 |
if not screenshot_bytes:
|
| 303 |
+
print("Failed to get screenshot after retries")
|
| 304 |
+
return -1.0, -1.0
|
| 305 |
|
| 306 |
current_screenshot_b64 = self.pil_to_base64(
|
| 307 |
image=Image.open(BytesIO(screenshot_bytes))
|
|
|
|
| 309 |
available_actions = self.controller.get_available_actions()
|
| 310 |
print(f"Available actions: {available_actions}")
|
| 311 |
|
| 312 |
+
# Get AI decision
|
| 313 |
+
if step == 1: # Final step - force guess
|
| 314 |
+
decision = self._get_final_guess(
|
| 315 |
+
history, current_screenshot_b64, available_actions
|
| 316 |
+
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 317 |
else:
|
|
|
|
| 318 |
decision = self.execute_agent_step(
|
| 319 |
history, step, current_screenshot_b64, available_actions
|
| 320 |
)
|
| 321 |
|
| 322 |
+
if not decision:
|
| 323 |
+
print("No decision from AI, using fallback")
|
| 324 |
+
decision = {
|
| 325 |
+
"reasoning": "AI decision failed",
|
| 326 |
+
"action_details": {
|
| 327 |
+
"action": "GUESS" if step == 1 else "PAN_RIGHT",
|
| 328 |
+
"lat": -1.0,
|
| 329 |
+
"lon": -1.0,
|
| 330 |
+
},
|
| 331 |
+
}
|
| 332 |
+
|
| 333 |
+
# UI callback
|
| 334 |
step_info = {
|
| 335 |
"step_num": step_num,
|
| 336 |
"max_steps": max_steps,
|
|
|
|
| 341 |
"is_final_step": step == 1,
|
| 342 |
"reasoning": decision.get("reasoning", "N/A"),
|
| 343 |
"action_details": decision.get("action_details", {"action": "N/A"}),
|
| 344 |
+
"history": history.copy(),
|
| 345 |
}
|
| 346 |
|
| 347 |
action_details = decision.get("action_details", {})
|
|
|
|
| 349 |
print(f"AI Reasoning: {decision.get('reasoning', 'N/A')}")
|
| 350 |
print(f"AI Action: {action}")
|
| 351 |
|
|
|
|
| 352 |
if step_callback:
|
| 353 |
try:
|
| 354 |
step_callback(step_info)
|
| 355 |
except Exception as e:
|
| 356 |
+
print(f"UI callback error: {e}")
|
| 357 |
|
| 358 |
+
# Add to history
|
| 359 |
self.add_step_to_history(history, current_screenshot_b64, decision)
|
| 360 |
|
| 361 |
# Execute action
|
| 362 |
if action == "GUESS":
|
| 363 |
+
lat = action_details.get("lat", -1.0)
|
| 364 |
+
lon = action_details.get("lon", -1.0)
|
| 365 |
+
print(f"Final guess: lat={lat}, lon={lon}")
|
| 366 |
+
|
| 367 |
+
# Validate coordinates
|
| 368 |
+
try:
|
| 369 |
+
lat_f, lon_f = float(lat), float(lon)
|
| 370 |
+
if -90 <= lat_f <= 90 and -180 <= lon_f <= 180:
|
| 371 |
+
return lat_f, lon_f
|
| 372 |
+
except (ValueError, TypeError):
|
| 373 |
+
pass
|
| 374 |
+
|
| 375 |
+
print("Invalid coordinates, returning error values")
|
| 376 |
+
return -1.0, -1.0
|
| 377 |
else:
|
| 378 |
self.execute_action(action)
|
| 379 |
|
| 380 |
+
print("Max steps reached without guess")
|
| 381 |
+
return -1.0, -1.0
|
| 382 |
+
|
| 383 |
+
def _get_final_guess(self, history, screenshot_b64, available_actions):
|
| 384 |
+
"""Get final guess from AI with simple retry."""
|
| 385 |
+
for retry in range(2):
|
| 386 |
+
try:
|
| 387 |
+
# If retry > 0, use a force prompt to ensure the AI returns a GUESS with coordinates.
|
| 388 |
+
if retry > 0:
|
| 389 |
+
history_text = self.generate_history_text(history)
|
| 390 |
+
force_prompt = f"""**FINAL STEP - MANDATORY GUESS**
|
| 391 |
+
You MUST return GUESS with coordinates. No other action allowed.
|
| 392 |
+
Remaining Steps: 1
|
| 393 |
+
Journey history: {history_text}
|
| 394 |
+
Provide your best lat/lon estimate based on all observed clues.
|
| 395 |
+
**MANDATORY JSON Format:**
|
| 396 |
+
{{"reasoning": "your analysis", "action_details": {{"action": "GUESS", "lat": 45.0, "lon": 2.0}} }}"""
|
| 397 |
+
|
| 398 |
+
message = self._create_message_with_history(
|
| 399 |
+
force_prompt, [screenshot_b64]
|
| 400 |
+
)
|
| 401 |
+
response = self.model.invoke(message)
|
| 402 |
+
decision = self._parse_agent_response(response)
|
| 403 |
+
else:
|
| 404 |
+
decision = self.execute_agent_step(
|
| 405 |
+
history, 1, screenshot_b64, available_actions
|
| 406 |
+
)
|
| 407 |
+
if (
|
| 408 |
+
decision
|
| 409 |
+
and decision.get("action_details", {}).get("action") == "GUESS"
|
| 410 |
+
):
|
| 411 |
+
return decision
|
| 412 |
+
print(f"AI didn't return GUESS, retry {retry + 1}/2")
|
| 413 |
+
except Exception as e:
|
| 414 |
+
print(f"AI call failed, retry {retry + 1}/2: {e}")
|
| 415 |
+
|
| 416 |
+
if retry == 0:
|
| 417 |
+
time.sleep(1)
|
| 418 |
+
|
| 419 |
+
# Fallback
|
| 420 |
+
return {
|
| 421 |
+
"reasoning": "AI failed to provide final guess after retries",
|
| 422 |
+
"action_details": {"action": "GUESS", "lat": -1.0, "lon": -1.0},
|
| 423 |
+
}
|
| 424 |
|
| 425 |
def analyze_image(self, image: Image.Image) -> Optional[Tuple[float, float]]:
|
| 426 |
image_b64 = self.pil_to_base64(image)
|