File size: 20,726 Bytes
3255c6b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
"""
Convert a single task folder with sequences into LIBERO-like demos with fields:

    actions, gripper_states, joint_states, robot_states, ee_states,
    agentview_images, eye_in_hand_images, agentview_depths, eye_in_hand_depths,
    agentview_segs, eye_in_hand_segs, agentview_boxes, eye_in_hand_boxes,
    rewards, dones

Expected layout (per task):
taskX/
  success/
    <seq_name>/
      camera_base.mp4        # agentview RGB
      camera_wrist.mp4       # eye-in-hand RGB
      trajectory.pkl         # dict-like (see below)
  masks/
    <seq_name>/
      masks/
        000000_id1.png, 000000_id2.png, 000001_id1.png, ...

We infer T (timesteps) from trajectory.pkl (preferred keys: robot_gripper_pose, timestamp).
We parse mask PNGs named "{frame:06d}_id{instance}.png" into a per-frame label map,
and compute per-frame boxes per instance id.

Trajectory .pkl keys (examples):
    ['robot_eef_pose', 'robot_eef_pose_vel', 'robot_joint', 'robot_joint_vel',
     'robot_gripper_pose', 'timestamp', 'task_description']

Actions policy:
- If 'robot_joint_vel' exists: actions = robot_joint_vel (T, DoF)
- Else if 'robot_eef_pose_vel' exists: actions = robot_eef_pose_vel (T, 6/7)
- Else: finite-difference of 'robot_joint' (pad last row with zeros).

Depth and eye-in-hand segs:
- If no depth available, we create zero arrays with the correct length and frame shape.
- If only one set of masks exists (agentview), we mirror it to eye-in-hand segs for compatibility.

Boxes:
- Stored in metainfo JSON as lists of [x1,y1,x2,y2] per frame (pixel coords).

Requires: numpy, opencv-python, h5py, pillow (PIL)
"""
import argparse, json, os, pickle, re, sys
from dataclasses import dataclass
from pathlib import Path
from typing import List, Tuple, Dict, Sequence, Optional, Any
import imageio

import numpy as np
import h5py
import cv2

from PIL import Image

MASK_RE = re.compile(r'^(?P<frame>\d+)_id(?P<inst>\d+)\.(?:png|jpg|jpeg|bmp)$', re.IGNORECASE)


# ---------- helpers ----------
def _ensure_uint8_rgb(img: np.ndarray) -> np.ndarray:
    arr = np.asarray(img)
    if arr.ndim == 2: arr = np.stack([arr]*3, axis=-1)
    if arr.shape[-1] == 4: arr = arr[..., :3]
    if arr.dtype != np.uint8:
        if np.issubdtype(arr.dtype, np.floating) and arr.max() <= 1.0:
            arr = (arr * 255.0 + 0.5).astype(np.uint8)
        else:
            arr = np.clip(arr, 0, 255).astype(np.uint8)
    return arr

def _label_to_color(label_map: np.ndarray,
                    color_map: Optional[Dict[int, Tuple[int,int,int]]] = None):
    H, W = label_map.shape
    colored = np.zeros((H, W, 3), dtype=np.uint8)
    color_map = {} if color_map is None else dict(color_map)
    for lid in np.unique(label_map):
        if lid == 0: continue
        if lid not in color_map:
            rng = np.random.RandomState(lid * 9973 % (2**31-1))
            color_map[lid] = tuple(int(x) for x in rng.randint(40, 220, size=3))
        colored[label_map == lid] = color_map[lid]
    return colored, color_map

def _overlay(rgb: np.ndarray, over_rgb: np.ndarray, alpha: float = 0.5) -> np.ndarray:
    out = (1.0 - alpha) * rgb.astype(np.float32) + alpha * over_rgb.astype(np.float32)
    return np.clip(out, 0, 255).astype(np.uint8)

def _draw_bboxes(rgb: np.ndarray,
                 bboxes: Sequence[Tuple[int, Sequence[int]]],
                 color_map: Optional[Dict[int, Tuple[int,int,int]]] = None) -> np.ndarray:
    img = rgb.copy()
    color_map = {} if color_map is None else color_map
    defined_labels = {'id40': 'bottle 1',
              'id20': 'bottle 2',
              'id60': 'bowl 1',
              'id100': 'robot',
              'id80': 'bowl 1'}
    for seg_id, box in bboxes:
        x, y, x2, y2 = [int(v) for v in box]
        if seg_id not in color_map:
            rng = np.random.RandomState(seg_id * 9973 % (2**31-1))
            color_map[seg_id] = tuple(int(x) for x in rng.randint(40, 220, size=3))
        bgr = color_map[seg_id][::-1]
        cv2.rectangle(img, (x, y), (x2, y2), bgr, 2)
        label = defined_labels[f"id{seg_id}"]
        (tw, th), _ = cv2.getTextSize(label, cv2.FONT_HERSHEY_SIMPLEX, 0.5, 1)
        cv2.rectangle(img, (x, y - th - 4), (x + tw + 4, y), bgr, -1)
        cv2.putText(img, label, (x + 2, y - 4),
                    cv2.FONT_HERSHEY_SIMPLEX, 0.5, (255,255,255), 1, cv2.LINE_AA)
    return img

# ---------- main ----------
def save_annotation_video_imageio(
    agentview_images: List[np.ndarray],
    agentview_segs: List[np.ndarray],
    agentview_bboxes: List[List[Tuple[int, Sequence[int]]]],
    out_path: str,
    fps: int = 20,
    resize: Optional[Tuple[int,int]] = None,
    seg_alpha: float = 0.5,
    layout: str = "hstack"
) -> str:
    """Save annotated rollout video with raw | bbox | seg-overlay panels using imageio."""
    assert len(agentview_images) == len(agentview_segs) == len(agentview_bboxes)
    T = len(agentview_images)
    if T == 0:
        raise ValueError("No frames to render")

    imgs = [_ensure_uint8_rgb(f) for f in agentview_images]
    segs = [np.asarray(s, dtype=np.int32) for s in agentview_segs]

    H, W = imgs[0].shape[:2]
    if resize is not None:
        W, H = resize
        imgs = [cv2.resize(im, (W, H), interpolation=cv2.INTER_LINEAR) for im in imgs]
        segs = [cv2.resize(s, (W, H), interpolation=cv2.INTER_NEAREST) for s in segs]
    else:
        imgs = [cv2.resize(im, (W, H), interpolation=cv2.INTER_LINEAR) if im.shape[:2] != (H, W) else im for im in imgs]
        segs = [cv2.resize(s, (W, H), interpolation=cv2.INTER_NEAREST) if s.shape != (H, W) else s for s in segs]

    color_map: Dict[int, Tuple[int,int,int]] = {}

    def compose(t: int) -> np.ndarray:
        raw = imgs[t]
        box_img = _draw_bboxes(raw, agentview_bboxes[t], color_map=color_map)
        seg_col, cm2 = _label_to_color(segs[t], color_map=color_map)
        color_map.update(cm2)
        seg_overlay = _overlay(raw, seg_col, alpha=seg_alpha)
        if layout == "hstack":
            return np.concatenate([raw, box_img, seg_overlay], axis=1)
        else:  # grid
            top = np.concatenate([raw, box_img], axis=1)
            bot = np.concatenate([seg_overlay, seg_col], axis=1)
            return np.concatenate([top, bot], axis=0)

    # --- Use imageio.get_writer ---
    with imageio.get_writer(out_path, fps=fps, codec="libx264") as writer:
        for t in range(T):
            frame = compose(t)
            writer.append_data(frame)  # frame must be (H,W,3) uint8

    return out_path

def natural_key(s: str):
    return [int(t) if t.isdigit() else t.lower() for t in re.split(r"(\d+)", s)]

def process_gripper_pose(robot_gripper_pose):
    raw = np.array(robot_gripper_pose)  # shape (T,)
    # binary states (open=1, closed=0)
    state = raw.astype(np.int32)

    # deltas using "previous" rule
    delta = np.zeros_like(state[:-1])
    prev = -1
    for t in range(0, len(state)-1):
        if state[t] != state[t+1]:
            delta[t] = 1 if state[t] < state[t+1] else -1
            prev = delta[t]
        else:
            delta[t] = prev  # carry forward previous action

    return delta

def process_video_rgb(path: Path) -> List[np.ndarray]:
    if cv2 is None:
        raise RuntimeError("OpenCV not available. Please install opencv-python.")
    cap = cv2.VideoCapture(str(path))
    if not cap.isOpened():
        raise RuntimeError(f"Cannot open video: {path}")
    frames = []
    while True:
        ok, frame = cap.read()
        if not ok:
            break
        # Resize to 256x256 and convert BGR->RGB
        frame = cv2.resize(frame, (256, 256), interpolation=cv2.INTER_LINEAR)
        frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
        frames.append(frame)
    cap.release()
    return frames

def parse_masks_dir(H, W, masks_dir: Path) -> Dict[int, Dict[int, np.ndarray]]:
    """
    Return nested dict: frame_idx -> {inst_id -> binary mask (H,W,1)}
    """
    out: Dict[int, Dict[int, np.ndarray]] = {}
    for f in sorted(masks_dir.iterdir(), key=lambda x: natural_key(x.name)):
        if not f.is_file(): continue
        m = MASK_RE.match(f.name)
        if not m: continue
        frame = int(m.group("frame"))
        inst = int(m.group("inst"))
        arr = np.array(Image.open(f).convert("L").resize((W, H)))  # (H,W) grayscale
        bin_mask = (arr > 0).astype(np.uint8)[..., None]  # (H,W,1)
        out.setdefault(frame, {})[inst] = bin_mask
    return out

def labelmap_and_boxes(H, W, per_inst: Dict[int, np.ndarray]) -> Tuple[np.ndarray, List[List[int]]]:
    """
    From {inst_id -> (H,W,1) mask}, build label map (H,W) with labels 30..K*30,
    and compute boxes as [x1,y1,x2,y2] for each instance (label>0), in order of inst_id.
    Returns (labelmap, boxes)
    """
    if not per_inst:
        return np.zeros((H,W,1), dtype=np.int32), []
    # Determine shape
    labelmap = np.zeros((H,W,1), dtype=np.int32)
    boxes: List[List[int]] = []
    # Sort instances for stable order
    for idx, inst_id in enumerate(sorted(per_inst.keys())):
        m = per_inst[inst_id][..., 0].astype(bool)
        label = (idx + 1)*20  # 0 reserved as background
        labelmap[m] = label
        # Bounding box
        ys, xs = np.where(m)
        if len(xs) == 0 or len(ys) == 0:
            pass
        else:
            x1, x2 = int(xs.min()), int(xs.max())
            y1, y2 = int(ys.min()), int(ys.max())
            boxes.append([label, [x1, y1, x2, y2]])
    return labelmap, boxes

def detect_noops_with_gripper_window(
    actions: np.ndarray,
    gripper_col: int = -1,
    tol: float = 1e-6,
    window: int = 6,
):
    """
    Return a boolean vector is_noop[T] where True marks a no-op step.
    A step is no-op if (a) all non-gripper dims are ~0 (|x|<tol), and
    (b) it's not within `window` frames after a gripper open/close change.

    Parameters
    ----------
    actions : (T, D) array
        Action vectors over time.
    gripper_col : int
        Index of the gripper signal column (default: last col).
    tol : float
        Tolerance to treat movement dims as zero.
    window : int
        Number of frames after a gripper state change to mark as active (non-noop).

    Returns
    -------
    is_noop : (T,) bool array
        True where the step is considered a no-op.
    active_gripper_window : (T,) bool array
        True where we are within the post-change window (non-noop region).
    """
    a = np.asarray(actions)
    assert a.ndim == 2 and a.shape[0] > 0, "actions must be (T, D)"
    T, D = a.shape

    # 1) movement no-op: all non-gripper dims are near zero
    if gripper_col < 0:
        g_idx = D + gripper_col
    else:
        g_idx = gripper_col
    assert 0 <= g_idx < D

    if D > 1:
        move = np.concatenate([a[:, :g_idx], a[:, g_idx+1:]], axis=1)
        movement_noop = np.all(np.abs(move) < tol, axis=1)
    else:
        movement_noop = np.ones(T, dtype=bool)  # only gripper present

    # 2) gripper activity window: detect state changes and mark window frames
    g = a[:, g_idx]

    # Convert to binary state: open=1, closed=0 (by sign/threshold)
    # Works for {-1,0,1} or continuous values (e.g., widths).
    state = (g > 0).astype(np.int8)

    # Change points where state flips
    changes = np.flatnonzero(np.diff(state, prepend=state[0]) != 0)

    active_gripper_window = np.zeros(T, dtype=bool)
    for t0 in changes:
        t1 = min(t0 + window, T)
        active_gripper_window[t0:t1] = True

    # Final no-op = movement_noop and NOT in gripper activity window
    is_noop = movement_noop & (~active_gripper_window)
    return is_noop, active_gripper_window

def process_sequence(seq_name: str, task_dir: Path, out_dir: Path, sequence_rename: Path):
    s_dir = task_dir / "success" / seq_name
    m_dir = task_dir / "masks" / seq_name / "masks"

    # --- Load trajectory ---
    pkl_path = s_dir / "trajectory.pkl"
    with open(pkl_path, "rb") as f:
        traj = pickle.load(f)
    
    task_description = traj['task_description'].lower().replace('.', '')
    T = len(traj['robot_eef_pose']) - 1

    delta_eef = traj['robot_eef_pose'][1:,:] - traj['robot_eef_pose'][:-1,:]
    delta_gripper = process_gripper_pose(traj['robot_gripper_pose'])
    delta_gripper = delta_gripper.reshape(T, 1)
    actions = np.concatenate([delta_eef, delta_gripper], axis=1)

    # --- Read videos as RGB ---
    base_vid = s_dir / "camera_base.mp4"
    agentview_images = process_video_rgb(base_vid)
    agentview_images = agentview_images[:T]
    H, W, _ = agentview_images[0].shape

    # --- Parse masks into label maps + boxes ---
    per_frame = parse_masks_dir(H, W, m_dir)
    agentview_segs = []
    agentview_bboxes = []

    for t, inst_dict in per_frame.items():
        if t >= T: continue
        labelmap, boxes = labelmap_and_boxes(H, W, inst_dict)
        if labelmap.size == 0:  # in case masks are missing
            continue
        if (labelmap.shape[0] != H) or (labelmap.shape[1] != W):
            # Resize nearest to match video shape
            labelmap = np.array(Image.fromarray(labelmap.astype(np.int32)).resize((W, H), resample=Image.NEAREST))
        agentview_segs.append(labelmap)
        agentview_bboxes.append(boxes)

    # save_annotation_video_imageio(
    #     agentview_images, agentview_segs, agentview_bboxes,
    #     out_path="annotations.mp4", fps=20, resize=(256,256)
    # )
    # 1/0
    # print(len(agentview_images))
    # print(len(agentview_segs))
    # print(len(agentview_bboxes))
    # print(len(actions))
    # print(actions); 
    # is_noop, active_win = detect_noops_with_gripper_window(actions, gripper_col=-1, tol=1e-5, window=6)
    data = {
        "episode_key": sequence_rename,
        "agentview_images": agentview_images,
        "agentview_segs":   agentview_segs,
        "agentview_boxes":  agentview_bboxes,
        "actions": actions,
        "task_description": task_description,
    }
    return data

def write_episode(
    out_dir: str,
    task_name: str,
    episode: Dict[str, Any],
):
    """
      {
        "episode_key": "20250711-13h_52m_58s",
        "agentview_images": [...],                 # list[(H,W,3) uint8]
        "agentview_segs":   [...],                 # list[(H,W) int]
        "agentview_boxes":  [...],                 # list[list[(id, [x,y,w,h])]]
        "actions": np.ndarray or None,             # (T,D)
        "task_description": "string",              # optional
      },
    """
    episode_key       = episode["episode_key"]
    h5_filename = f"{task_name}_{episode_key}.hdf5"
    meta_filename = f"{task_name}_{episode_key}_metainfo.json"

    h5_path   = os.path.join(out_dir, h5_filename)
    meta_path = os.path.join(out_dir, meta_filename)

    # Load or start metainfo (single JSON for all episodes)
    if os.path.exists(meta_path):
        with open(meta_path, "r") as f:
            metainfo = json.load(f)
    else:
        metainfo = {task_name: {}}

    with h5py.File(h5_path, "a") as f:  # append if file already exists
        root = f.require_group("data")
        ep = episode
        episode_key       = ep["episode_key"]
        agentview_images  = ep["agentview_images"]
        agentview_segs    = ep["agentview_segs"]
        agentview_boxes   = ep["agentview_boxes"]
        actions           = ep.get("actions", None)
        task_description  = ep.get("task_description", "")

        # --- lengths & alignment ---
        lens = [len(agentview_images), len(agentview_segs), len(agentview_boxes)]
        if actions is not None: lens.append(len(actions))
        T = min(l for l in lens if l > 0)
        assert T > 0, f"[{episode_key}] nothing to write"

        agentview_images = agentview_images[:T]
        agentview_segs   = agentview_segs[:T]
        agentview_boxes  = agentview_boxes[:T]
        if actions is None:
            actions = np.zeros((T, 1), dtype=np.float32)
        else:
            actions = np.asarray(actions)[:T]

        # --- stack visuals ---
        agentview_rgb = np.stack(agentview_images, axis=0)  # (T,H,W,3)
        agentview_seg = np.stack([np.asarray(s, dtype=np.int32) for s in agentview_segs], axis=0)  # (T,H,W)
        _, H, W, _ = agentview_seg.shape

        # --- placeholders for missing streams/states ---
        eye_in_hand_rgb   = np.zeros_like(agentview_rgb, dtype=np.uint8)
        agentview_depth   = np.zeros((T, H, W), dtype=np.float32)
        eye_in_hand_depth = np.zeros((T, H, W), dtype=np.float32)
        eye_in_hand_seg   = np.zeros((T, H, W), dtype=np.int32)

        gripper_states = np.zeros((T, 1), dtype=np.float32)
        joint_states   = np.zeros((T, 0), dtype=np.float32)
        ee_states      = np.zeros((T, 6), dtype=np.float32)   # [pos(3), ori(3)]
        robot_states   = np.zeros((T, 0), dtype=np.float32)

        dones   = np.zeros(T, dtype=np.uint8);   dones[-1] = 1
        rewards = np.zeros(T, dtype=np.uint8); rewards[-1] = 1

        # --- create / overwrite episode group ---
        if episode_key in root:
            del root[episode_key]  # clean if re-writing
        ep_grp  = root.create_group(episode_key)
        obs_grp = ep_grp.create_group("obs")

        # states
        obs_grp.create_dataset("gripper_states", data=gripper_states)
        obs_grp.create_dataset("joint_states",   data=joint_states)
        obs_grp.create_dataset("ee_states",      data=ee_states)
        obs_grp.create_dataset("ee_pos",         data=ee_states[:, :3])
        obs_grp.create_dataset("ee_ori",         data=ee_states[:, 3:])

        # visuals
        obs_grp.create_dataset("agentview_rgb",     data=agentview_rgb)
        obs_grp.create_dataset("eye_in_hand_rgb",   data=eye_in_hand_rgb)
        obs_grp.create_dataset("agentview_depth",   data=agentview_depth)
        obs_grp.create_dataset("eye_in_hand_depth", data=eye_in_hand_depth)
        obs_grp.create_dataset("agentview_seg",     data=agentview_seg)
        obs_grp.create_dataset("eye_in_hand_seg",   data=eye_in_hand_seg)

        # top-level (episode)
        ep_grp.create_dataset("actions",      data=actions)
        ep_grp.create_dataset("robot_states", data=robot_states)
        ep_grp.create_dataset("rewards",      data=rewards)
        ep_grp.create_dataset("dones",        data=dones)

        # --- update metainfo JSON for this episode ---
        if task_name not in metainfo:
            metainfo[task_name] = {}
        if episode_key not in metainfo[task_name]:
            metainfo[task_name][episode_key] = {}

        metainfo[task_name][episode_key].update({
            "success": True,
            "initial_state": robot_states[0].tolist() if len(robot_states) else [],
            "task_nouns": [],                         # fill if you want
            "task_description": task_description,
            "exo_boxes": agentview_boxes,            # per-frame boxes you provided
            "ego_boxes": [[] for _ in range(T)],     # none available
        })

    # write/merge metainfo once at the end
    with open(meta_path, "w") as f:
        json.dump(metainfo, f, indent=2)

    return {"hdf5": h5_path, "metainfo": meta_path}

def main():
    p = argparse.ArgumentParser(description="Convert sequences to LIBERO-like demos.")
    p.add_argument("--task_dir", type=str, help="Path to task folder (contains success/ and masks/).")
    p.add_argument("--out_root", type=str, required=True, help="Target directory where <task_name>/<task_name>_<seq>.hdf5 is written.")
    args = p.parse_args()

    task_dir = Path(args.task_dir).expanduser().resolve()
    task_name = task_dir.name
    out_root = Path(args.out_root).expanduser().resolve() 
    out_root.mkdir(parents=True, exist_ok=True)

    success_dir = task_dir / "success"
    masks_dir = task_dir / "masks"
    if not success_dir.is_dir() or not masks_dir.is_dir():
        print("[ERROR] task_dir must contain 'success/' and 'masks/'")
        sys.exit(1)

    success_seqs = {d.name for d in success_dir.iterdir() if d.is_dir()}
    mask_seqs = {d.name for d in masks_dir.iterdir() if d.is_dir()}
    seqs = sorted(list(success_seqs & mask_seqs), key=natural_key)

    results = []
    from tqdm import tqdm
    for i, name in tqdm(enumerate(seqs)):
        info = process_sequence(name, task_dir, out_root, sequence_rename=f'demo_{i+1}')
        
        write_episode(
            out_dir=args.out_root,
            task_name=info['task_description'],
            episode=info,
        )

    # # Write a small manifest JSON
    # manifest = {"task": task_name, "outputs": results}
    # (out_root / f"{task_name}_manifest.json").write_text(json.dumps(manifest, indent=2))
    # print(f"[DONE] Manifest saved to {out_root / (task_name + '_manifest.json')}")

if __name__ == "__main__":
    main()