Dom/rpa_vision_v3
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: VLM grounding direct (Qwen2.5-VL) — nouvelle stratégie de résolution

Browse Source 
Nouvelle approche basée sur les recherches état de l'art :
- _resolve_by_grounding() : le VLM retourne directement les coordonnées
  (pas de SomEngine + numérotation intermédiaire)
- Utilise Qwen2.5-VL (entraîné pour le GUI grounding) au lieu de qwen3-vl
- Parse les formats natifs : bbox_2d, JSON x/y, arrays bruts
- Fallback multi-image : screenshot + crop → grounding sans description
- Identification des icônes via Qwen2.5-VL (meilleur que qwen3-vl)

Résultats sur session réelle (validation locale) :
- Éléments avec texte (Word, Document, Fichier) : 100% corrects
- Icônes sans texte (Windows logo, disquette) : en cours d'amélioration

Cascade strict mode :
0. Grounding VLM direct (Qwen2.5-VL) — NOUVEAU
0.5. Template matching pour icônes
1. VLM Quick Find (fallback)
1.5. SoM + VLM
2. Template matching strict

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
Dom
2026-03-31 18:55:00 +02:00
parent 875367dea9
commit d99b17394a
2 changed files with 243 additions and 25 deletions
Download Patch File Download Diff File Expand all files Collapse all files

230
agent_v0/server_v1/api_stream.py
View File

@@ -3366,6 +3366,206 @@ def _vlm_quick_find(
return None
# ---------------------------------------------------------------------------
# Résolution par VLM Grounding Direct (Qwen2.5-VL)
# ---------------------------------------------------------------------------
def _resolve_by_grounding(
screenshot_path: str,
target_spec: Dict[str, Any],
screen_width: int,
screen_height: int,
) -> Optional[Dict[str, Any]]:
"""Résoudre une cible via grounding VLM direct (Qwen2.5-VL).
Le VLM reçoit le screenshot + une description textuelle et retourne
directement les coordonnées (bbox_2d) de l'élément. Pas de SomEngine,
pas de numérotation — le VLM est entraîné pour le grounding UI.
Approche plus fiable que SomEngine+VLM pour les icônes et éléments
visuels sans texte (logo Windows, disquette, bouton fermer).
"""
import base64
import io
import re
t0 = time.time()
# Construire la description de la cible
by_text = target_spec.get("by_text", "").strip()
vlm_desc = target_spec.get("vlm_description", "").strip()
window_title = target_spec.get("window_title", "").strip()
if by_text:
description = by_text
elif vlm_desc:
description = vlm_desc
else:
return None
# Redimensionner le screenshot (800px de large pour le VLM)
try:
from PIL import Image as PILImage
img = PILImage.open(screenshot_path)
orig_w, orig_h = img.size
target_w = 800
ratio = target_w / orig_w
img_small = img.resize((target_w, int(orig_h * ratio)))
small_w, small_h = img_small.size
buf = io.BytesIO()
img_small.save(buf, format="JPEG", quality=75)
shot_b64 = base64.b64encode(buf.getvalue()).decode()
except Exception as e:
logger.warning("Grounding : erreur redimensionnement — %s", e)
return None
# Construire le prompt — Qwen2.5-VL retourne naturellement des bbox_2d
prompt = (
f"Look at this screenshot. Find: {description}\n"
"Where is it? Give the center position as percentage of the image.\n"
'Answer ONLY with JSON: {"x": 0.XX, "y": 0.YY}'
)
# Appel VLM (Qwen2.5-VL pour le grounding)
try:
import requests as _requests
resp = _requests.post("http://localhost:11434/api/chat", json={
"model": "qwen2.5vl:7b",
"messages": [
{"role": "system", "content": "You locate UI elements on screenshots. Return coordinates."},
{"role": "user", "content": prompt, "images": [shot_b64]},
],
"stream": False,
"options": {"temperature": 0.1, "num_predict": 80},
}, timeout=60)
content = resp.json().get("message", {}).get("content", "")
except Exception as e:
logger.info("Grounding VLM timeout/erreur : %s", e)
return None
elapsed = time.time() - t0
# Parser la réponse — Qwen2.5-VL retourne soit bbox_2d en pixels, soit JSON %
x_pct, y_pct = None, None
# Format 1 : bbox_2d en pixels [x, y] ou [x1, y1, x2, y2]
bbox_match = re.search(r'"bbox_2d"\s*:\s*\[([^\]]+)\]', content)
if bbox_match:
coords = [float(v.strip()) for v in bbox_match.group(1).split(",")]
if len(coords) == 2:
x_pct = coords[0] / small_w
y_pct = coords[1] / small_h
elif len(coords) >= 4:
x_pct = (coords[0] + coords[2]) / 2 / small_w
y_pct = (coords[1] + coords[3]) / 2 / small_h
# Format 2 : JSON {"x": 0.XX, "y": 0.YY}
if x_pct is None:
json_match = re.search(r'"x"\s*:\s*([\d.]+).*?"y"\s*:\s*([\d.]+)', content)
if json_match:
x_val, y_val = float(json_match.group(1)), float(json_match.group(2))
# Si > 1, c'est en pixels
if x_val > 1:
x_pct = x_val / small_w
y_pct = y_val / small_h
else:
x_pct = x_val
y_pct = y_val
# Format 3 : {"x_pct": 0.XX, "y_pct": 0.YY}
if x_pct is None:
pct_match = re.search(r'"x_pct"\s*:\s*([\d.]+).*?"y_pct"\s*:\s*([\d.]+)', content)
if pct_match:
x_pct = float(pct_match.group(1))
y_pct = float(pct_match.group(2))
# Format 4 : array brut [x1, y1, x2, y2] ou [x, y]
if x_pct is None:
arr_match = re.search(r'\[[\s]*([\d.]+)\s*,\s*([\d.]+)(?:\s*,\s*([\d.]+)\s*,\s*([\d.]+))?\s*\]', content)
if arr_match:
vals = [float(v) for v in arr_match.groups() if v is not None]
if len(vals) >= 4:
x_pct = (vals[0] + vals[2]) / 2 / small_w
y_pct = (vals[1] + vals[3]) / 2 / small_h
elif len(vals) == 2:
x_pct = vals[0] / small_w
y_pct = vals[1] / small_h
if x_pct is None or y_pct is None:
# Fallback multi-image : screenshot + crop → grounding sans description
anchor_b64 = target_spec.get("anchor_image_base64", "")
if anchor_b64:
try:
prompt_mi = (
"Image 1 is a screenshot. Image 2 shows a UI element.\n"
"Find where Image 2 appears on Image 1.\n"
'Return position: {"x": NNN, "y": NNN} in pixels of Image 1.'
)
resp2 = _requests.post("http://localhost:11434/api/chat", json={
"model": "qwen2.5vl:7b",
"messages": [
{"role": "user", "content": prompt_mi, "images": [shot_b64, anchor_b64]},
],
"stream": False,
"options": {"temperature": 0.1, "num_predict": 50},
}, timeout=60)
content2 = resp2.json().get("message", {}).get("content", "")
elapsed = time.time() - t0
# Parser tous les formats
arr2 = re.search(r'\[[\s]*([\d.]+)\s*,\s*([\d.]+)(?:\s*,\s*([\d.]+)\s*,\s*([\d.]+))?\s*\]', content2)
if arr2:
vals = [float(v) for v in arr2.groups() if v is not None]
if len(vals) >= 4:
x_pct = (vals[0] + vals[2]) / 2 / small_w
y_pct = (vals[1] + vals[3]) / 2 / small_h
elif len(vals) == 2:
x_pct = vals[0] / small_w
y_pct = vals[1] / small_h
if x_pct is None:
json2 = re.search(r'"x"\s*:\s*([\d.]+).*?"y"\s*:\s*([\d.]+)', content2)
if json2:
x_pct = float(json2.group(1)) / small_w
y_pct = float(json2.group(2)) / small_h
if x_pct is not None:
logger.info("Grounding multi-image OK (%.1fs)", elapsed)
except Exception as e:
logger.debug("Grounding multi-image erreur: %s", e)
if x_pct is None or y_pct is None:
logger.info(
"Grounding : réponse non parsable (%.1fs) — %s",
elapsed, content[:120],
)
return None
# Valider les bornes
if not (0.0 <= x_pct <= 1.0 and 0.0 <= y_pct <= 1.0):
logger.info("Grounding : coordonnées hors bornes (%.3f, %.3f)", x_pct, y_pct)
return None
logger.info(
"Grounding OK [qwen2.5vl] : '%s' → (%.4f, %.4f) en %.1fs",
description[:50], x_pct, y_pct, elapsed,
)
return {
"resolved": True,
"method": "grounding_vlm",
"x_pct": round(x_pct, 6),
"y_pct": round(y_pct, 6),
"matched_element": {
"label": description[:60],
"type": "grounding",
"role": "grounding_vlm",
"confidence": 0.85,
},
"score": 0.85,
}
# ---------------------------------------------------------------------------
# Résolution Set-of-Mark : SomEngine (détection) + VLM (identification)
# ---------------------------------------------------------------------------
@@ -3770,9 +3970,29 @@ def _resolve_target_sync(
vlm_description = _build_target_description(target_spec)
# ---------------------------------------------------------------
# Étape 0 : Template matching PRIORITAIRE pour les icônes sans texte
# Les crops 80x80 sont très discriminants pour les icônes (logo Windows,
# disquette, croix). Le VLM se trompe souvent sur ces éléments.
# Étape 0 : Grounding VLM Direct (Qwen2.5-VL)
# Le VLM reçoit le screenshot + description textuelle et retourne
# directement les coordonnées. Plus fiable que SomEngine + numérotation.
# ---------------------------------------------------------------
grounding_desc = by_text_strict or vlm_description
if grounding_desc:
grounding_result = _resolve_by_grounding(
screenshot_path=screenshot_path,
target_spec=target_spec,
screen_width=screen_width,
screen_height=screen_height,
)
if grounding_result and grounding_result.get("resolved"):
logger.info(
"Strict resolve GROUNDING : OK (%.4f, %.4f) pour '%s'",
grounding_result.get("x_pct", 0),
grounding_result.get("y_pct", 0),
grounding_desc[:50],
)
return grounding_result
# ---------------------------------------------------------------
# Étape 0.5 : Template matching pour icônes sans texte (crop 80x80)
# ---------------------------------------------------------------
if not by_text_strict:
result = _resolve_by_template_matching(
@@ -3784,13 +4004,13 @@ def _resolve_target_sync(
)
if result and result.get("score", 0) >= 0.70:
logger.info(
"Strict resolve icon : template matching OK (score=%.3f) pour icône sans texte",
"Strict resolve TEMPLATE : icon match (score=%.3f)",
result.get("score", 0),
)
return result
# ---------------------------------------------------------------
# Étape 1 : VLM Quick Find (compréhension sémantique)
# Étape 1 : VLM Quick Find (fallback, multi-image)
# ---------------------------------------------------------------
if vlm_description or anchor_image_b64:
vlm_result = _vlm_quick_find(