app.py

# Access: https://BinKhoaLe1812-Sall-eGarbageDetection.hf.space/ui

# ───────────────────────── app.py (Sall-e demo) ─────────────────────────
# FastAPI ▸ upload image ▸ multi-model garbage detection ▸ ADE-20K
# semantic segmentation (Water / Garbage) ▸ A* navigation ▸ H.264 video
# =======================================================================

import os, uuid, threading, shutil, time, heapq, cv2, numpy as np
from PIL import Image
import uvicorn
from fastapi import FastAPI, File, UploadFile, Request
from fastapi.responses import HTMLResponse, StreamingResponse, Response
from fastapi.staticfiles import StaticFiles


# ── Vision libs ─────────────────────────────────────────────────────────
import torch, yolov5, ffmpeg
from ultralytics import YOLO
from transformers import (
    DetrImageProcessor, DetrForObjectDetection,
    SegformerFeatureExtractor, SegformerForSemanticSegmentation
)
# from sklearn.neighbors import NearestNeighbors
from inference_sdk import InferenceHTTPClient


# ── Folders / files ─────────────────────────────────────────────────────
BASE        = "/home/user/app"
CACHE       = f"{BASE}/cache"
UPLOAD_DIR  = f"{CACHE}/uploads"
OUTPUT_DIR  = f"{BASE}/outputs"
MODEL_DIR   = f"{BASE}/model"
SPRITE      = f"{BASE}/sprite.png"

os.makedirs(UPLOAD_DIR, exist_ok=True)
os.makedirs(OUTPUT_DIR, exist_ok=True)
os.makedirs(CACHE     , exist_ok=True)
os.environ["TRANSFORMERS_CACHE"] = CACHE
os.environ["HF_HOME"]           = CACHE

# ── Load models once  ───────────────────────────────────────────────────
print("🔄  Loading models …")
model_self  = YOLO(f"{MODEL_DIR}/garbage_detector.pt")                 # YOLOv11(l)
model_yolo5 = yolov5.load(f"{MODEL_DIR}/yolov5-detect-trash-classification.pt")
processor_detr = DetrImageProcessor.from_pretrained(f"{MODEL_DIR}/detr")
model_detr     = DetrForObjectDetection.from_pretrained(f"{MODEL_DIR}/detr")
feat_extractor = SegformerFeatureExtractor.from_pretrained(
                    "nvidia/segformer-b4-finetuned-ade-512-512")
segformer      = SegformerForSemanticSegmentation.from_pretrained(
                    "nvidia/segformer-b4-finetuned-ade-512-512")
model_animal = YOLO(f"{MODEL_DIR}/yolov8n.pt") # Load COCO pre-trained YOLOv8 for animal detection
print("✅  Models ready\n")

# ── ADE-20K palette + custom mapping (verbatim) ─────────────────────────
# ADE20K palette
ade_palette = np.array([
    [0, 0, 0], [120, 120, 120], [180, 120, 120], [6, 230, 230], [80, 50, 50],
    [4, 200, 3], [120, 120, 80], [140, 140, 140], [204, 5, 255], [230, 230, 230],
    [4, 250, 7], [224, 5, 255], [235, 255, 7], [150, 5, 61], [120, 120, 70],
    [8, 255, 51], [255, 6, 82], [143, 255, 140], [204, 255, 4], [255, 51, 7],
    [204, 70, 3], [0, 102, 200], [61, 230, 250], [255, 6, 51], [11, 102, 255],
    [255, 7, 71], [255, 9, 224], [9, 7, 230], [220, 220, 220],
    [112, 9, 255], [8, 255, 214], [7, 255, 224], [255, 184, 6], [10, 255, 71],
    [255, 41, 10], [7, 255, 255], [224, 255, 8], [102, 8, 255], [255, 61, 6],
    [255, 194, 7], [255, 122, 8], [0, 255, 20], [255, 8, 41], [255, 5, 153],
    [6, 51, 255], [235, 12, 255], [160, 150, 20], [0, 163, 255], [140, 140, 140],
    [250, 10, 15], [20, 255, 0], [31, 255, 0], [255, 31, 0], [255, 224, 0],
    [153, 255, 0], [0, 0, 255], [255, 71, 0], [0, 235, 255], [0, 173, 255],
    [31, 0, 255], [11, 200, 200], [255, 82, 0], [0, 255, 245], [0, 61, 255],
    [0, 255, 112], [0, 255, 133], [255, 0, 0], [255, 163, 0], [255, 102, 0],
    [194, 255, 0], [0, 143, 255], [51, 255, 0], [0, 82, 255], [0, 255, 41],
    [255, 0, 255], [255, 0, 245], [255, 0, 102], [255, 173, 0], [255, 0, 20],
    [255, 184, 184], [0, 31, 255], [0, 255, 61], [0, 71, 255], [255, 0, 204],
    [0, 255, 194], [0, 255, 82], [0, 10, 255], [0, 112, 255], [51, 0, 255],
    [0, 194, 255], [0, 122, 255], [0, 255, 163], [255, 153, 0], [0, 255, 10],
    [255, 112, 0], [143, 255, 0], [82, 0, 255], [163, 255, 0], [255, 235, 0],
    [8, 184, 170], [133, 0, 255], [0, 255, 92], [184, 0, 255], [255, 0, 31],
    [0, 184, 255], [0, 214, 255], [255, 0, 112], [92, 255, 0], [0, 224, 255],
    [112, 224, 255], [70, 184, 160], [163, 0, 255], [153, 0, 255], [71, 255, 0],
    [255, 0, 163], [255, 204, 0], [255, 0, 143], [0, 255, 235], [133, 255, 0],
    [255, 0, 235], [245, 0, 255], [255, 0, 122], [255, 245, 0], [10, 190, 212],
    [214, 255, 0], [0, 204, 255], [20, 0, 255], [255, 255, 0], [0, 153, 255],
    [0, 41, 255], [0, 255, 204], [41, 0, 255], [41, 255, 0], [173, 0, 255],
    [0, 245, 255], [71, 0, 255], [122, 0, 255], [0, 255, 184], [0, 92, 255],
    [184, 255, 0], [0, 133, 255], [255, 214, 0], [25, 194, 194], [102, 255, 0],
    [92, 0, 255]
], dtype=np.uint8)

custom_class_map = {
    "Garbage":                 [(255, 8, 41), (235, 255, 7), (255, 5, 153), (255, 0, 102)],
    "Water":                   [(0, 102, 200), (11, 102, 255), (31, 0, 255), (10, 0, 255), (9, 7, 230)],
    "Grass / Vegetation":      [(10, 255, 71), (143, 255, 140)],
    "Tree / Natural Obstacle": [(4, 200, 3), (235, 12, 255), (255, 6, 82), (255, 163, 0)],
    "Sand / Soil / Ground":    [(80, 50, 50), (230, 230, 230)],
    "Buildings / Structures":  [(255, 0, 255), (184, 0, 255), (120, 120, 120), (7, 255, 224)],
    "Sky / Background":        [(180, 120, 120)],
    "Undetecable":             [(0, 0, 0)],
    "Unknown Class": []
}
TOL = 30  # RGB tolerance

# Segment class [150, 5, 61] is only detectable as garbage if it's large enough 
def interpret_rgb_class(decoded_img):
    ambiguous_rgb = np.array([150, 5, 61])
    matches = np.all(np.abs(decoded_img - ambiguous_rgb) <= TOL, axis=-1)
    match_ratio = np.count_nonzero(matches) / matches.size
    return "garbage" if match_ratio > 0.15 else "sand"

# Masking zones (Garbage and Water zone to be travelable)
def build_masks(seg):
    """
    Returns three binary masks at (H,W):
    water_mask   – 1 = water
    garbage_mask – 1 = semantic “Garbage” pixels
    movable_mask – union of water & garbage (robot can travel here)
    """
    decoded = ade_palette[seg]
    water_mask   = np.zeros(seg.shape, np.uint8)
    garbage_mask = np.zeros_like(water_mask)
    # Resolve ambiguity: (150,5,61) → Sand or Garbage?
    context_label = interpret_rgb_class(decoded)
    resolved_map = custom_class_map.copy()
    # Dynamically re-assign the ambiguous RGB class
    if context_label == "garbage":
        resolved_map["Garbage"].append((150, 5, 61))
        resolved_map["Sand / Soil / Ground"] = [rgb for rgb in resolved_map["Sand / Soil / Ground"] if rgb != (150, 5, 61)]
    else: # Fall back as appointed to be sth else
        resolved_map["Sand / Soil / Ground"].append((150, 5, 61))
        resolved_map["Garbage"] = [rgb for rgb in resolved_map["Garbage"] if rgb != (150, 5, 61)]
    # Append water pixels to water_mask
    for rgb in custom_class_map["Water"]:
        water_mask |= (np.abs(decoded - rgb).max(axis=-1) <= TOL)
    # Append gb pixels to garbage_mask
    for rgb in custom_class_map["Garbage"]:
        garbage_mask |= (np.abs(decoded - rgb).max(axis=-1) <= TOL)
    movable_mask = water_mask | garbage_mask
    return water_mask, garbage_mask, movable_mask

# Garbage mask can be highlighted in red
def highlight_chunk_masks_on_frame(frame, labels, objs, color_uncollected=(0, 0, 128), color_collected=(0, 128, 0), alpha=0.8):
    """
    Overlays semi-transparent colored regions for garbage chunks on the frame.
    `objs` must have 'pos' and 'col' keys. The collection status changes the overlay color.
    """
    overlay = frame.copy()
    for i, obj in enumerate(objs):
        x, y = obj["pos"]
        lab = labels[y, x]
        if lab == 0:
            continue
        mask = (labels == lab).astype(np.uint8)
        contours, _ = cv2.findContours(mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
        color = color_collected if obj["col"] else color_uncollected
        # drawContours on overlay
        cv2.drawContours(overlay, contours, -1, color, thickness=cv2.FILLED)
    # Blend overlay with original frame using alpha
    return cv2.addWeighted(overlay, alpha, frame, 1 - alpha, 0)
# Water mask to be blue
def highlight_water_mask_on_frame(frame, binary_mask, color=(255, 0, 0), alpha=0.3):
    """
    Overlays semi-transparent colored mask (binary) on the frame.
    """
    overlay = frame.copy()
    mask = binary_mask.astype(np.uint8) * 255
    contours, _ = cv2.findContours(mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
    # drawContours on overlay
    cv2.drawContours(overlay, contours, -1, color, thickness=cv2.FILLED)
    return cv2.addWeighted(overlay, alpha, frame, 1 - alpha, 0)


# ── A* and KNN over binary water grid ─────────────────────────────────
def astar(start, goal, occ):
    h = lambda a, b: abs(a[0] - b[0]) + abs(a[1] - b[1])
    N8 = [(-1,-1), (-1,0), (-1,1), (0,-1), (0,1), (1,-1), (1,0), (1,1)]
    openq = [(0, start)]
    g = {start: 0}
    came = {}
    visited = set()
    H, W = occ.shape
    while openq:
        _, cur = heapq.heappop(openq)
        if cur == goal:
            p = [cur]
            while cur in came:
                cur = came[cur]
                p.append(cur)
            return p[::-1]
        if cur in visited:
            continue
        visited.add(cur)
        for dx, dy in N8:
            nx, ny = cur[0] + dx, cur[1] + dy
            if not (0 <= nx < W and 0 <= ny < H): continue
            if occ[ny, nx] == 0: continue
            # prevent corner-cutting on diagonals
            if abs(dx) == 1 and abs(dy) == 1:
                if occ[cur[1] + dy, cur[0]] == 0 or occ[cur[1], cur[0] + dx] == 0:
                    continue
            next_node = (nx, ny)
            ng = g[cur] + 1
            if next_node not in g or ng < g[next_node]:
                g[next_node] = ng
                f = ng + h(next_node, goal)
                heapq.heappush(openq, (f, next_node))
                came[next_node] = cur
    print(f"[A*] No path from {start} to {goal}")
    return []

# KNN fit optimal path
def knn_path(start, targets, occ):
    todo = targets[:]; path=[]
    cur  = tuple(start)
    reachable = []; unreachable = []
    for t in todo:
        seg = astar(cur, tuple(t), occ)
        if seg:
            if path and path[-1] == seg[0]:
                seg = seg[1:]
            path.extend(seg)
            reachable.append(t)
            cur = tuple(t)
        else:
            unreachable.append(t)
            print(f"⚠️ Unreachable garbage at {t}")
    return path, reachable


# ── Robot sprite/class -──────────────────────────────────────────────────
class Robot:
    def __init__(self, sprite, speed=2000): # Declare the robot's physical stats and routing (position, speed, movement, path)
        img = Image.open(sprite).convert("RGBA").resize((40, 40))
        self.png = np.array(img)
        if self.png.shape[-1] != 4:
            raise ValueError("Sprite image must have 4 channels (RGBA)")
        self.png = np.array(Image.open(sprite).convert("RGBA").resize((40,40)))
        self.speed = speed
        self.pos = [20, 20]  # Fallback spawn with body offset at top-left
    def step(self, path):
        while path:
            dx, dy = path[0][0] - self.pos[0], path[0][1] - self.pos[1]
            dist = (dx * dx + dy * dy) ** 0.5
            if dist <= self.speed:
                self.pos = list(path.pop(0))
            else: # If valid path within
                r = self.speed / dist
                new_x = self.pos[0] + dx * r
                new_y = self.pos[1] + dy * r
                # Clip to valid region with 20px margin (for body offset)
                self.pos = [
                    int(np.clip(new_x, 20, 640 - 20)),
                    int(np.clip(new_y, 20, 640 - 20))
                ]
            # Break after one logical move to avoid overshooting
            break


# ── Static-web ────────────────────────────────────────────────────────── 
from fastapi.responses import JSONResponse, FileResponse
from fastapi.middleware.cors import CORSMiddleware
app = FastAPI()
app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_methods=["*"],
    allow_headers=["*"],
)
app.mount("/statics", StaticFiles(directory="statics"), name="statics")
video_ready={}
@app.get("/ui", response_class=HTMLResponse)
async def serve_index():
    p = "statics/index.html"
    if os.path.exists(p):
        print("[STATIC] Serving index.html")
        return FileResponse(p)
    print("[STATIC] index.html not found")
    return JSONResponse(status_code=404, content={"detail":"Not found"})
def _uid(): return uuid.uuid4().hex[:8]


# ── End-points ──────────────────────────────────────────────────────────
# User upload environment img here
@app.post("/upload/")
async def upload(file:UploadFile=File(...)):
    uid=_uid(); dest=f"{UPLOAD_DIR}/{uid}_{file.filename}"
    with open(dest,"wb") as bf: shutil.copyfileobj(file.file,bf)
    threading.Thread(target=_pipeline, args=(uid,dest)).start()
    return {"user_id":uid}

# Health check, make sure the video generator is alive and debug which video id is processed (multiple video can be processed at 1 worker)
@app.get("/check_video/{uid}")
def chk(uid:str): return {"ready":video_ready.get(uid,False)}

# Where the final video being saved
@app.get("/video/{uid}")
def stream(uid:str):
    vid=f"{OUTPUT_DIR}/{uid}.mp4"
    if not os.path.exists(vid): return Response(status_code=404)
    return StreamingResponse(open(vid,"rb"), media_type="video/mp4")

# ─── Detect animal/wildlife ─────────────────────────────────────────────────
# Init clients
# https://universe.roboflow.com/team-hope-mmcyy/hydroquest | https://universe.roboflow.com/sky-sd2zq/bird_only-pt0bm/model/1
import base64, requests
def roboflow_infer(image_path, api_url, api_key):
    with open(image_path, "rb") as image_file:
        files = {"file": image_file}
        res = requests.post(
            f"{api_url}?api_key={api_key}&confidence=70",  # Add threshold to URL
            files=files
        )
    print(f"[Roboflow] {res.status_code} response")
    try:
        return res.json()
    except Exception as e:
        print("[Roboflow JSON decode error]", e)
        return {}

# Animal detection endpoint (animal, fish, bird as target classes)
@app.post("/animal/")
async def detect_animals(file: UploadFile = File(...)):
    img_id = _uid()
    img_path = f"{UPLOAD_DIR}/{img_id}_{file.filename}"
    with open(img_path, "wb") as f:
        shutil.copyfileobj(file.file, f)
    print(f"[Animal] Uploaded image: {img_path}")
    # Read and prepare detection
    image = cv2.imread(img_path)
    detections = []

    # 1. YOLOv8 local
    print("[Animal] Detecting via YOLOv8…")
    try:
        results = model_animal(image)[0]
        for box in results.boxes:
            conf = box.conf[0].item()
            if conf >= 0.70:
                cls_id = int(box.cls[0].item())
                label = model_animal.names[cls_id].lower()
                if label in ["dog", "cat", "cow", "horse", "elephant", "bear", "zebra", "giraffe", "bird"]:
                    x1, y1, x2, y2 = map(int, box.xyxy[0].tolist())
                    detections.append(((x1, y1, x2, y2), f"Animal Alert {conf}"))
    except Exception as e:
        print("[YOLOv8 Error]", e)

    # Hide on production-level
    print("[API] Roboflow key:", os.getenv("ROBOFLOW_KEY", "❌ not set"))

    # 2. Roboflow Fish
    try:
        print("[Animal] Detecting via Roboflow Fish model…")
        fish_response = roboflow_infer(
            img_path,
            "https://detect.roboflow.com/hydroquest/1",
            api_key=os.getenv("ROBOFLOW_KEY", "")
        )
        for pred in fish_response.get("predictions", []):
            if pred["confidence"] >= 0.70:
                acc = pred["confidence"]
                x1 = int(pred["x"] - pred["width"] / 2)
                y1 = int(pred["y"] - pred["height"] / 2)
                x2 = int(pred["x"] + pred["width"] / 2)
                y2 = int(pred["y"] + pred["height"] / 2)
                detections.append(((x1, y1, x2, y2), f"Fish Alert {acc}"))
        print("[Roboflow Fish Response]", fish_response)
    except Exception as e:
        print("[Roboflow Fish Error]", e)

    # 3. Roboflow Bird
    try:
        print("[Animal] Detecting via Roboflow Bird model…")
        bird_response = roboflow_infer(
            img_path,
            "https://detect.roboflow.com/bird_only-pt0bm/1",
            api_key=os.getenv("ROBOFLOW_KEY", "")
        )
        for pred in bird_response.get("predictions", []):
            if pred["confidence"] >= 0.70:
                acc = pred["confidence"]
                x1 = int(pred["x"] - pred["width"] / 2)
                y1 = int(pred["y"] - pred["height"] / 2)
                x2 = int(pred["x"] + pred["width"] / 2)
                y2 = int(pred["y"] + pred["height"] / 2)
                detections.append(((x1, y1, x2, y2), f"Bird Alert {acc}"))
        print("[Roboflow Bird Response]", bird_response)
    except Exception as e:
        print("[Roboflow Bird Error]", e)

    # Count detection
    print(f"[Animal] Total detections: {len(detections)}")
    # Write label
    for (x1, y1, x2, y2), label in detections:
        cv2.rectangle(image, (x1, y1), (x2, y2), (0, 0, 255), 2)
        cv2.putText(image, label, (x1, y1 - 10), cv2.FONT_HERSHEY_SIMPLEX, 0.8, (0, 0, 255), 2)
    # Write img
    result_path = f"{OUTPUT_DIR}/{img_id}_animal.jpg"
    cv2.imwrite(result_path, image)
    return FileResponse(result_path, media_type="image/jpeg")


# ── Core pipeline (runs in background thread) ───────────────────────────
def _pipeline(uid,img_path):
    print(f"▶️ [{uid}] processing")
    bgr=cv2.resize(cv2.imread(img_path),(640,640)); rgb=cv2.cvtColor(bgr,cv2.COLOR_BGR2RGB)
    pil=Image.fromarray(rgb)

    # 1- Segmentation → masking each segmented zone with pytorch
    with torch.no_grad():
        inputs = feat_extractor(pil, return_tensors="pt")
        seg_logits = segformer(**inputs).logits
    # Tensor run by CPU
    seg_tensor = seg_logits.argmax(1)[0].cpu()
    if seg_tensor.numel() == 0:
        print(f"❌ [{uid}] segmentation failed (empty tensor)")
        video_ready[uid] = True
        return
    # Resize the tensor to 640x640
    seg = cv2.resize(seg_tensor.numpy(), (640, 640), interpolation=cv2.INTER_NEAREST)
    print(f"🧪 [{uid}] segmentation input shape: {inputs['pixel_values'].shape}")
    water_mask, garbage_mask, movable_mask = build_masks(seg) # movable zone = water and garbage masks

    # 2- Garbage detection (3 models) → keep centres on water 
    detections=[]
    # Detect garbage chunks (from segmentation)
    num_cc, labels = cv2.connectedComponents(garbage_mask.astype(np.uint8))
    chunk_centres = []
    for lab in range(1, num_cc):
        ys, xs = np.where(labels == lab)
        if xs.size == 0: # safety
            continue
        chunk_centres.append([int(xs.mean()), int(ys.mean())])
    print(f"🧠 {len(chunk_centres)} garbage chunk detected")
    # Detect garbage object by within travelable zones
    for r in model_self(bgr):                      # YOLOv11 (self-trained)
        detections += [b.xyxy[0].tolist() for b in r.boxes]
    r = model_yolo5(bgr)                           # YOLOv5
    if hasattr(r, 'pred') and len(r.pred) > 0:
        detections += [p[:4].tolist() for p in r.pred[0]]
    inp=processor_detr(images=pil,return_tensors="pt")
    with torch.no_grad(): out=model_detr(**inp)    # DETR
    post = processor_detr.post_process_object_detection(
        outputs=out,
        target_sizes=torch.tensor([pil.size[::-1]]),
        threshold=0.5
    )[0]
    detections += [b.tolist() for b in post["boxes"]]
    # centre & mask filter (the garbage lies within travelable zone are collectable)
    centres = []
    for x1, y1, x2, y2 in detections: # Define IoU heuristic
        '''
        We conduct a 20% allowance whether the center 
        of the detected garbage's bbox lies within the travelable zone
        which was segmented earlier to be the water and garbage zone
        '''
        x1, y1, x2, y2 = map(int, [x1, y1, x2, y2])
        x1 = max(0, min(x1, 639)); y1 = max(0, min(y1, 639))
        x2 = max(0, min(x2, 639)); y2 = max(0, min(y2, 639))
        box_mask = movable_mask[y1:y2, x1:x2]              # ← switch to movable_mask
        if box_mask.size == 0:
            continue
        if np.count_nonzero(box_mask) / box_mask.size >= 0.2:
            centres.append([int((x1 + x2) / 2), int((y1 + y2) / 2)])
    # add chunk centres and deduplicate
    centres.extend(chunk_centres)
    centres = [list(c) for c in {tuple(c) for c in centres}]
    if not centres: # No garbages within travelable zone
        print(f"🛑 [{uid}] no reachable garbage"); video_ready[uid]=True; return
    else: # Garbage within valid travelable zone
        print(f"🧠 {len(centres)} garbage objects on water selected from {len(detections)} detections")

    # 3- Robot initialization, position and navigation
    # find all (y,x) within water-zone
    ys, xs = np.where(water_mask)
    if len(ys)==0:
        # no travelable zone → bail out
        print(f"❌ [{uid}] no water to spawn on")
        video_ready[uid] = True
        return
    # sort by y, then x
    idx = np.lexsort((xs, ys))
    spawn_y, spawn_x = int(ys[idx[0]]), int(xs[idx[0]])
    # enforce 20px margin so sprite never pokes out
    spawn_x = np.clip(spawn_x,   20, 640-20)
    spawn_y = np.clip(spawn_y,   20, 640-20)
    robot = Robot(SPRITE)
    # Robot will be spawn on the closest movable mask to top-left
    robot.pos = [spawn_x, spawn_y]
    path, unreachable_targets = knn_path(robot.pos, centres, movable_mask)
    objs = [{"pos": p, "col": False, "unreachable": False} for p in centres if p not in unreachable_targets]
    objs += [{"pos": p, "col": False, "unreachable": True} for p in unreachable_targets]


    # 4- Video synthesis
    out_tmp=f"{OUTPUT_DIR}/{uid}_tmp.mp4"
    vw=cv2.VideoWriter(out_tmp,cv2.VideoWriter_fourcc(*"mp4v"),10.0,(640,640))
    objs=[{"pos":p,"col":False} for p in centres]
    bg = bgr.copy()
    for _ in range(15000): # safety frames
        frame=bg.copy()
        # Draw garbage chunk masks in red-to-green (semi-transparent)
        frame = highlight_chunk_masks_on_frame(frame, labels, objs)                               # 🔴 garbage overlay
        frame = highlight_water_mask_on_frame(frame, water_mask)  # 🔵 water overlay
        # Draw object detections as red (to green) dots
        for o in objs:
            color = (0, 0, 128) if not o["col"] else (0, 128, 0)
            x, y = o["pos"]
            cv2.circle(frame, (x, y), 6, color, -1)
        # Robot displacement
        robot.step(path)
        sp = robot.png
        sprite_h, sprite_w = sp.shape[:2]
        rx, ry = robot.pos
        x1, y1 = rx - sprite_w // 2, ry - sprite_h // 2
        x2, y2 = x1 + sprite_w, y1 + sprite_h
        # Clip boundaries to image size
        x1_clip, x2_clip = max(0, x1), min(frame.shape[1], x2)
        y1_clip, y2_clip = max(0, y1), min(frame.shape[0], y2)
        # Adjust sprite crop accordingly
        sx1, sy1 = x1_clip - x1, y1_clip - y1
        sx2, sy2 = sprite_w - (x2 - x2_clip), sprite_h - (y2 - y2_clip)
        sprite_crop = sp[sy1:sy2, sx1:sx2]
        alpha = sprite_crop[:, :, 3] / 255.0
        alpha = np.stack([alpha] * 3, axis=-1)
        bgroi = frame[y1_clip:y2_clip, x1_clip:x2_clip]
        blended = (alpha * sprite_crop[:, :, :3] + (1 - alpha) * bgroi).astype(np.uint8)
        frame[y1_clip:y2_clip, x1_clip:x2_clip] = blended
        # collection check
        for o in objs:
            if not o["col"] and np.hypot(o["pos"][0]-robot.pos[0], o["pos"][1]-robot.pos[1]) <= 20:
                o["col"]=True
        vw.write(frame)
        if all(o["col"] for o in objs): break
        if not path: break
    vw.release()

    # 5- Convert to H.264
    final=f"{OUTPUT_DIR}/{uid}.mp4"
    ffmpeg.input(out_tmp).output(final,vcodec="libx264",pix_fmt="yuv420p").run(overwrite_output=True,quiet=True)
    os.remove(out_tmp); video_ready[uid]=True
    print(f"✅ [{uid}] video ready → {final}")


# ── Run locally (HF Space ignores since built with Docker image) ────────
if __name__=="__main__":
    uvicorn.run(app,host="0.0.0.0",port=7860)