Datasets:

zhicao
/

openpi

"""
Final DROID Preprocessing Script
Preprocesses DROID episodes for ControlNet+UNet training with:
1. 7 mesh vertices (computed per frame as ground truth)
2. CoTracker on 1000 points including the 7 mesh vertices (similar to LIBERO's 1098)
3. No manual chunking - uses CoTracker's automatic handling
4. Saves in This&That RLDS-compatible format with visibility masks
For each episode:
- Exterior view: 1000 tracked points (first 7 are mesh vertices, remaining 993 are arm-shaped)
- Wrist view: 1000 tracked points (300 sparse + 700 dense bottom 60%-100%)
- Ground truth mesh vertices saved separately for comparison
- Original 180×320 resolution
- Visibility masks saved for occlusion handling
"""

import sys
from pathlib import Path
sys.path.append(str(Path(__file__).parent.parent))

import os
import numpy as np
import torch
import mediapy as media
import tensorflow as tf
tf.config.set_visible_devices([], 'GPU')
import tensorflow_datasets as tfds
import cv2
import datetime
import re
import json
from tqdm import tqdm
from scipy.spatial.transform import Rotation as R

from utils.load_camera_calibration import CameraCalibrationLoader
from utils.franka_mesh_projection import FrankaMeshProjector


# 7 gripper offsets in gripper frame (before rotation)
GRIPPER_OFFSETS = np.array([
    [0.0, 0.0, 0.0],      # 0: gripper base
    [0.0, 0.045, 0.161],  # 1: finger 1 tip
    [0.0, -0.045, 0.161], # 2: finger 2 tip
    [0.0, 0.045, 0.13],   # 3: finger 1 end
    [0.0, -0.045, 0.13],  # 4: finger 2 end
    [0.0, 0.0, 0.13],     # 5: gripper center front
    [0.0, 0.0, 0.065],    # 6: gripper center middle
])


def euler_xyz_to_rotation_matrix(euler_xyz):
    """Convert Euler XYZ angles to rotation matrix."""
    return R.from_euler('xyz', euler_xyz).as_matrix()


def transform_gripper_offsets(action):
    """
    Transform gripper offsets using action position and rotation.
    Args:
        action: [x, y, z, rx, ry, rz, gripper] - Euler XYZ rotation
    Returns:
        gripper_points_3d: [7, 3] array of 3D points in world frame
    """
    pos = action[:3]
    rot_euler = action[3:6]
    rot_matrix = euler_xyz_to_rotation_matrix(rot_euler)
    gripper_points_3d = (rot_matrix @ GRIPPER_OFFSETS.T).T + pos
    return gripper_points_3d


def sample_arm_shaped_points(mesh_2d_visible, img_h, img_w, num_points=993, seed=None):
    """
    Sample points in arm shape around visible mesh vertices.
    Strategy:
    - Many points per visible mesh vertex (Gaussian, σ=15 pixels)
    - Points along lines connecting mesh vertices
    - Remaining points uniform random
    Args:
        mesh_2d_visible: [N_visible, 2] visible mesh vertex projections
        img_h, img_w: Image dimensions
        num_points: Total points to sample (default 993 for 1000 total with 7 mesh)
        seed: Random seed for reproducibility
    Returns:
        points: [num_points, 2] sampled points in pixel coordinates
    """
    if seed is not None:
        np.random.seed(seed)

    points = []
    num_visible = len(mesh_2d_visible)

    if num_visible == 0:
        # No mesh visible, return uniform random
        return np.random.rand(num_points, 2) * [img_w, img_h]

    # 1. Gaussian around each visible mesh vertex (15 per mesh)
    points_per_mesh = min(15, num_points // num_visible)
    gaussian_sigma = 15.0  # pixels

    for mesh_pt in mesh_2d_visible:
        gaussian_pts = np.random.randn(points_per_mesh, 2) * gaussian_sigma + mesh_pt
        # Clip to image bounds
        gaussian_pts[:, 0] = np.clip(gaussian_pts[:, 0], 0, img_w - 1)
        gaussian_pts[:, 1] = np.clip(gaussian_pts[:, 1], 0, img_h - 1)
        points.append(gaussian_pts)

    # 2. Lines between visible meshes (if multiple visible)
    if num_visible >= 2:
        # Sample along lines connecting consecutive visible points
        points_per_line = 6
        for i in range(num_visible - 1):
            line_pts = np.linspace(mesh_2d_visible[i], mesh_2d_visible[i+1], points_per_line + 2)
            points.append(line_pts[1:-1])  # Exclude endpoints (already in Gaussian)

    # 3. Fill remaining with uniform random
    current_count = sum(len(p) for p in points)
    remaining = num_points - current_count

    if remaining > 0:
        uniform_pts = np.random.rand(remaining, 2) * [img_w, img_h]
        points.append(uniform_pts)

    # Concatenate and ensure exact count
    all_points = np.vstack(points) if points else np.empty((0, 2))

    if len(all_points) < num_points:
        # Pad with uniform random
        extra = np.random.rand(num_points - len(all_points), 2) * [img_w, img_h]
        all_points = np.vstack([all_points, extra])
    elif len(all_points) > num_points:
        # Truncate
        all_points = all_points[:num_points]

    return all_points


def get_wrist_gripper_mesh_2d(gripper_state, img_h=180, img_w=320):
    """
    Generate 7-point 2D gripper mesh for wrist camera (INITIAL TRACKING QUERIES).
    Similar to exterior camera's 3D projected mesh, these are the initial query
    points for CoTracker to track. The gripper is at ~70% from left (not centered).
    Args:
        gripper_state: float in [-1, 1] from action[6]
                      -1 = fully closed, +1 = fully open
        img_h, img_w: Image dimensions (default DROID native 180x320)
    Returns:
        mesh_2d: [7, 2] array with (x, y) pixel coordinates for initial tracking
                 0: gripper base
                 1: right finger tip
                 2: left finger tip
                 3: right finger joint
                 4: left finger joint
                 5: center front
                 6: center back
    """
    # Manually annotated keypoints (same as fixed mesh)
    # OPEN state (gripper_state = 0.0)
    mesh_open = np.array([
        [141, 108],  # 0: palm base
        [253, 106],  # 1: RIGHT finger tip
        [129, 131],  # 2: LEFT finger tip
        [290, 130],  # 3: RIGHT finger joint
        [168, 156],  # 4: LEFT finger joint
        [284, 156],  # 5: right palm interior
        [95, 138],   # 6: left palm interior
    ], dtype=np.float32)

    # CLOSED state (gripper_state = 1.0)
    mesh_closed = np.array([
        [190, 102],  # 0: palm base
        [198, 102],  # 1: RIGHT finger tip
        [195, 120],  # 2: LEFT finger tip
        [216, 119],  # 3: RIGHT finger joint
        [200, 152],  # 4: LEFT finger joint
        [245, 150],  # 5: right palm interior
        [168, 126],  # 6: left palm interior
    ], dtype=np.float32)

    # Interpolate between OPEN (gs=0) and CLOSED (gs=1)
    alpha = gripper_state  # 0 = use mesh_open, 1 = use mesh_closed
    mesh_2d = (1 - alpha) * mesh_open + alpha * mesh_closed

    return mesh_2d


def get_wrist_fixed_mesh_2d(gripper_state, img_h=180, img_w=320):
    """
    Generate 7-point FIXED 2D mesh that moves with gripper state (NOT TRACKED).
    Uses manually annotated keypoints and interpolates between them based on gripper state.
    IMPORTANT: Gripper state convention is INVERTED in DROID dataset:
        gripper_state = 0.0 → OPEN (wide finger spacing)
        gripper_state = 1.0 → CLOSED (narrow finger spacing)
    Args:
        gripper_state: float in [0, 1] from action[6]
                      0.0 = fully OPEN, 1.0 = fully CLOSED
        img_h, img_w: Image dimensions (default DROID native 180x320)
    Returns:
        mesh_2d: [7, 2] array with (x, y) pixel coordinates
                 0: palm base
                 1: RIGHT finger tip
                 2: LEFT finger tip
                 3: RIGHT finger joint
                 4: LEFT finger joint
                 5: right palm interior
                 6: left palm interior
    """
    # Manually annotated keypoints from web annotation tool
    # Convention: gripper_state 0.0=OPEN, 1.0=CLOSED

    # OPEN state (gripper_state = 0.0) - wide finger spacing
    mesh_open = np.array([
        [141, 108],  # 0: palm base
        [253, 106],  # 1: RIGHT finger tip
        [129, 131],  # 2: LEFT finger tip
        [290, 130],  # 3: RIGHT finger joint
        [168, 156],  # 4: LEFT finger joint
        [284, 156],  # 5: right palm interior
        [95, 138],   # 6: left palm interior
    ], dtype=np.float32)

    # CLOSED state (gripper_state = 1.0) - narrow finger spacing
    mesh_closed = np.array([
        [190, 102],  # 0: palm base
        [198, 102],  # 1: RIGHT finger tip (only 8px from left!)
        [195, 120],  # 2: LEFT finger tip
        [216, 119],  # 3: RIGHT finger joint
        [200, 152],  # 4: LEFT finger joint
        [245, 150],  # 5: right palm interior
        [168, 126],  # 6: left palm interior
    ], dtype=np.float32)

    # Interpolate between OPEN (gs=0) and CLOSED (gs=1)
    # gripper_state goes from 0 (open) to 1 (closed)
    alpha = gripper_state  # 0 = use mesh_open, 1 = use mesh_closed
    mesh_2d = (1 - alpha) * mesh_open + alpha * mesh_closed

    return mesh_2d


def sample_double_grid(n, img_h, img_w):
    """
    Sample two offset grids for background coverage (matches LIBERO preprocessing).
    Creates 2×(n×n) grid points across the image for stable background tracking.
    Args:
        n: Grid size (e.g., n=7 gives 7×7×2=98 points)
        img_h, img_w: Image dimensions
    Returns:
        points: [2*n*n, 2] array of (x, y) coordinates
    """
    # Grid 1: from 5% to 85%
    u1 = np.linspace(0.05 * img_w, 0.85 * img_w, n)
    v1 = np.linspace(0.05 * img_h, 0.85 * img_h, n)
    u1, v1 = np.meshgrid(u1, v1)
    grid1 = np.stack([u1.flatten(), v1.flatten()], axis=-1)

    # Grid 2: from 15% to 95% (offset from grid1)
    u2 = np.linspace(0.15 * img_w, 0.95 * img_w, n)
    v2 = np.linspace(0.15 * img_h, 0.95 * img_h, n)
    u2, v2 = np.meshgrid(u2, v2)
    grid2 = np.stack([u2.flatten(), v2.flatten()], axis=-1)

    # Combine both grids
    points = np.vstack([grid1, grid2]).astype(np.float32)
    return points


def sample_uniform_grid(
    n,
    img_h,
    img_w,
    pad_ratio=0.05,
    crop_left=None,
    crop_right=None,
    crop_top=0.0,
    crop_bottom=None,
    out_w=None,
    out_h=None,
):
    """Sample an n×n grid; optionally uniform after crop+resize transform."""
    if crop_left is None or crop_right is None:
        u = np.linspace(pad_ratio * img_w, (1.0 - pad_ratio) * img_w, n)
        v = np.linspace(pad_ratio * img_h, (1.0 - pad_ratio) * img_h, n)
    else:
        if crop_bottom is None:
            crop_bottom = float(img_h)
        crop_w = float(crop_right - crop_left)
        crop_h = float(crop_bottom - crop_top)
        if out_w is None:
            out_w = crop_w
        if out_h is None:
            out_h = crop_h
        u_out = np.linspace(pad_ratio * out_w, (1.0 - pad_ratio) * (out_w - 1.0), n)
        v_out = np.linspace(pad_ratio * out_h, (1.0 - pad_ratio) * (out_h - 1.0), n)
        u = u_out * (crop_w / float(out_w)) + float(crop_left)
        v = v_out * (crop_h / float(out_h)) + float(crop_top)
    uu, vv = np.meshgrid(u, v)
    return np.stack([uu.flatten(), vv.flatten()], axis=-1).astype(np.float32)


def sample_wrist_points(img_h, img_w, num_sparse=300, num_dense=700, seed=None):
    """
    Sample points for wrist view: sparse uniform + dense in bottom region.
    Bottom region: Y-coords from 60% to 100% of image height (bottom 40%)
    This is where the gripper typically appears in wrist camera view.
    Args:
        img_h, img_w: Image dimensions
        num_sparse: Number of sparse uniform points (default 300)
        num_dense: Number of dense points in bottom region (default 700)
        seed: Random seed
    Returns:
        points: [num_sparse + num_dense, 2] sampled points (1000 total by default)
    """
    if seed is not None:
        np.random.seed(seed)

    # Sparse uniform across full image
    sparse = np.random.rand(num_sparse, 2) * [img_w, img_h]

    # Dense in bottom 60%-100% region
    # For 180 height: 60% = 108, 100% = 180
    y_min = int(img_h * 0.60)
    y_max = img_h
    y_range = y_max - y_min

    dense_x = np.random.rand(num_dense) * img_w
    dense_y = np.random.rand(num_dense) * y_range + y_min
    dense = np.column_stack([dense_x, dense_y])

    return np.vstack([sparse, dense])


def create_temporal_queries(points_2d, T, num_mesh=7):
    """
    Create CoTracker queries with temporal sampling.
    Mesh vertices (first 7 points) always start at frame 0 (ground truth).
    Remaining points are sampled from random frames across the video.
    Args:
        points_2d: [N, 2] array of (x, y) spatial coordinates
        T: Number of frames in video
        num_mesh: Number of mesh points that start at t=0 (default 7)
    Returns:
        queries: [N, 3] array with (t, x, y) for CoTracker
    """
    N = len(points_2d)
    queries = np.zeros((N, 3), dtype=np.float32)

    # Mesh vertices always start at frame 0 (ground truth from projection)
    queries[:num_mesh, 0] = 0
    queries[:num_mesh, 1:] = points_2d[:num_mesh]

    # Remaining points: sample from random frames
    if T > 1:
        random_frames = np.random.randint(0, T, size=N - num_mesh)
    else:
        random_frames = np.zeros(N - num_mesh, dtype=np.int32)

    queries[num_mesh:, 0] = random_frames
    queries[num_mesh:, 1:] = points_2d[num_mesh:]

    return queries


def track_with_variance_filter(cotracker, video_tensor, queries_tensor, device,
                                var_threshold=10.0, preserve_indices=None):
    """
    Track points with variance filtering and forward/backward tracking.
    Performs both forward and backward tracking passes, averages results,
    then filters out low-variance (static/noisy) points. Mesh points are
    always preserved. If too many points are filtered, resamples with jitter.
    Args:
        cotracker: CoTracker model
        video_tensor: [1, T, 3, H, W] video tensor
        queries_tensor: [1, N, 3] query tensor (t, x, y)
        device: torch device
        var_threshold: Minimum variance to keep point (default 10.0, matches LIBERO preprocessing)
        preserve_indices: List/array of point indices to always preserve (mesh + grid points)
    Returns:
        tracks: [T, N, 2] numpy array
        vis: [T, N] numpy array
    """
    B, T, C, H, W = video_tensor.shape
    N = queries_tensor.shape[1]

    # Forward tracking
    with torch.no_grad():
        tracks_fwd, vis_fwd = cotracker(
            video_tensor,
            queries=queries_tensor,
            backward_tracking=False
        )

    # Backward tracking
    with torch.no_grad():
        tracks_bwd, vis_bwd = cotracker(
            video_tensor,
            queries=queries_tensor,
            backward_tracking=True
        )

    # Average forward and backward for robustness
    tracks = (tracks_fwd + tracks_bwd) / 2.0
    vis = (vis_fwd + vis_bwd) / 2.0

    # Compute variance per point (measure of motion/dynamics)
    # High variance = dynamic point, Low variance = static/noisy
    variance = torch.var(tracks[0], dim=0).sum(dim=-1)  # [N]

    # Preserve specified indices (mesh + grid points, always keep them)
    if preserve_indices is not None:
        variance[preserve_indices] = float('inf')  # Ensures they're never filtered

    # Filter: keep points with variance above threshold
    valid_idx = torch.where(variance > var_threshold)[0]

    # If we filtered too many points, resample from valid ones with jitter
    if len(valid_idx) < N:
        n_needed = N - len(valid_idx)

        # Resample from valid points
        resample_idx = valid_idx[torch.randint(len(valid_idx), (n_needed,), device=device)]
        new_queries = queries_tensor[:, resample_idx].clone()

        # Add spatial jitter to resampled queries (5% of image height)
        noise = torch.randn_like(new_queries[:, :, 1:]) * 0.05 * H
        new_queries[:, :, 1:] = torch.clamp(new_queries[:, :, 1:] + noise, 0, None)

        # Re-track resampled points (backward only for efficiency)
        with torch.no_grad():
            new_tracks, new_vis = cotracker(
                video_tensor,
                queries=new_queries,
                backward_tracking=True
            )

        # Concatenate valid + resampled tracks
        tracks = torch.cat([tracks[:, :, valid_idx], new_tracks], dim=2)
        vis = torch.cat([vis[:, :, valid_idx], new_vis], dim=2)
    else:
        # Keep only valid points
        tracks = tracks[:, :, valid_idx]
        vis = vis[:, :, valid_idx]

    return tracks[0].cpu().numpy(), vis[0].cpu().numpy()


def track_with_variance_filter_batched(cotracker, video_tensor, queries_tensor, device,
                                        var_threshold=10.0, preserve_indices=None, batch_size=600):
    """
    Batched version of track_with_variance_filter for OOM handling.
    Splits queries into batches to reduce memory usage. Uses same logic as
    track_with_variance_filter but processes points in smaller chunks.
    Args:
        cotracker: CoTracker model
        video_tensor: [1, T, 3, H, W] video tensor
        queries_tensor: [1, N, 3] query tensor (t, x, y)
        device: torch device
        var_threshold: Minimum variance to keep point (default 10.0)
        preserve_indices: List/array of point indices to always preserve
        batch_size: Number of points per batch (default 500)
    Returns:
        tracks: [T, N, 2] numpy array
        vis: [T, N] numpy array
    """
    B, T, C, H, W = video_tensor.shape
    N = queries_tensor.shape[1]

    # Calculate number of batches
    num_batches = (N + batch_size - 1) // batch_size
    print(f"  [BATCHED] Tracking {N} points in {num_batches} batches of {batch_size}")

    # Forward tracking in batches
    tracks_fwd_list = []
    vis_fwd_list = []

    for batch_idx in range(num_batches):
        start_idx = batch_idx * batch_size
        end_idx = min(start_idx + batch_size, N)
        batch_queries = queries_tensor[:, start_idx:end_idx, :]

        with torch.no_grad():
            batch_tracks_fwd, batch_vis_fwd = cotracker(
                video_tensor,
                queries=batch_queries,
                backward_tracking=False
            )

        tracks_fwd_list.append(batch_tracks_fwd)
        vis_fwd_list.append(batch_vis_fwd)

        # Free GPU memory after each batch
        torch.cuda.empty_cache()

    # Concatenate forward tracking results
    tracks_fwd = torch.cat(tracks_fwd_list, dim=2)  # [1, T, N, 2]
    vis_fwd = torch.cat(vis_fwd_list, dim=2)  # [1, T, N]

    # Backward tracking in batches
    tracks_bwd_list = []
    vis_bwd_list = []

    for batch_idx in range(num_batches):
        start_idx = batch_idx * batch_size
        end_idx = min(start_idx + batch_size, N)
        batch_queries = queries_tensor[:, start_idx:end_idx, :]

        with torch.no_grad():
            batch_tracks_bwd, batch_vis_bwd = cotracker(
                video_tensor,
                queries=batch_queries,
                backward_tracking=True
            )

        tracks_bwd_list.append(batch_tracks_bwd)
        vis_bwd_list.append(batch_vis_bwd)

        # Free GPU memory after each batch
        torch.cuda.empty_cache()

    # Concatenate backward tracking results
    tracks_bwd = torch.cat(tracks_bwd_list, dim=2)  # [1, T, N, 2]
    vis_bwd = torch.cat(vis_bwd_list, dim=2)  # [1, T, N]

    # Average forward and backward for robustness
    tracks = (tracks_fwd + tracks_bwd) / 2.0
    vis = (vis_fwd + vis_bwd) / 2.0

    # Compute variance per point (measure of motion/dynamics)
    variance = torch.var(tracks[0], dim=0).sum(dim=-1)  # [N]

    # Preserve specified indices (mesh + grid points, always keep them)
    if preserve_indices is not None:
        variance[preserve_indices] = float('inf')

    # Filter: keep points with variance above threshold
    valid_idx = torch.where(variance > var_threshold)[0]

    # If we filtered too many points, resample from valid ones with jitter
    if len(valid_idx) < N:
        n_needed = N - len(valid_idx)

        # Resample from valid points
        resample_idx = valid_idx[torch.randint(len(valid_idx), (n_needed,), device=device)]
        new_queries = queries_tensor[:, resample_idx].clone()

        # Add spatial jitter to resampled queries (5% of image height)
        noise = torch.randn_like(new_queries[:, :, 1:]) * 0.05 * H
        new_queries[:, :, 1:] = torch.clamp(new_queries[:, :, 1:] + noise, 0, None)

        # Re-track resampled points (backward only for efficiency)
        # Also batch this if needed
        if n_needed > batch_size:
            new_tracks_list = []
            new_vis_list = []

            num_resample_batches = (n_needed + batch_size - 1) // batch_size
            for batch_idx in range(num_resample_batches):
                start_idx = batch_idx * batch_size
                end_idx = min(start_idx + batch_size, n_needed)
                batch_new_queries = new_queries[:, start_idx:end_idx, :]

                with torch.no_grad():
                    batch_new_tracks, batch_new_vis = cotracker(
                        video_tensor,
                        queries=batch_new_queries,
                        backward_tracking=True
                    )

                new_tracks_list.append(batch_new_tracks)
                new_vis_list.append(batch_new_vis)
                torch.cuda.empty_cache()

            new_tracks = torch.cat(new_tracks_list, dim=2)
            new_vis = torch.cat(new_vis_list, dim=2)
        else:
            with torch.no_grad():
                new_tracks, new_vis = cotracker(
                    video_tensor,
                    queries=new_queries,
                    backward_tracking=True
                )

        # Concatenate valid + resampled tracks
        tracks = torch.cat([tracks[:, :, valid_idx], new_tracks], dim=2)
        vis = torch.cat([vis[:, :, valid_idx], new_vis], dim=2)
    else:
        # Keep only valid points
        tracks = tracks[:, :, valid_idx]
        vis = vis[:, :, valid_idx]

    return tracks[0].cpu().numpy(), vis[0].cpu().numpy()


def load_cotracker(device=None):
    """
    Load CoTracker v3 offline model.
    Args:
        device: torch device to load model on. If None, uses cuda if available.
    Returns:
        model: CoTracker model
        device: Device model is loaded on
    """
    from cotracker.predictor import CoTrackerPredictor

    if device is None:
        device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

    # Try multiple checkpoint locations
    cotracker_paths = [
        '/mnt/kevin/vlm_models/cotracker/scaled_offline.pth',
        '/mnt/kevin/vlm_models/hub/checkpoints/scaled_offline.pth',
        '/root/workspace/code/wmrl/Dual-Dynamics-Models/DROID-main/co-tracker/checkpoints/scaled_offline.pth',
    ]

    cotracker_checkpoint = None
    for path in cotracker_paths:
        if Path(path).exists():
            cotracker_checkpoint = path
            print(f"Found CoTracker checkpoint: {cotracker_checkpoint}")
            break

    if cotracker_checkpoint is None:
        raise FileNotFoundError(f"CoTracker checkpoint not found. Tried:\n" + "\n".join(cotracker_paths))

    model = CoTrackerPredictor(checkpoint=cotracker_checkpoint)
    model = model.to(device)
    model.eval()
    return model, device


def find_closest_calibration(episode, uuid_list):
    """Find closest calibration by timestamp (FIXED UUID MATCHING)."""
    try:
        # Extract timestamp from file_path
        file_path = episode['episode_metadata']['file_path'].numpy().decode('utf-8')
        # Path format: .../YYYY-MM-DD/Day_Month_DD_HH:MM:SS_YYYY/...
        date_match = re.search(r'/(\d{4})-(\d{2})-(\d{2})/[^/]+_(\d{1,2}):(\d{2}):(\d{2})_\d{4}/', file_path)

        if not date_match:
            return None

        year, month, day, hour, minute, second = date_match.groups()
        episode_ts = datetime.datetime(int(year), int(month), int(day), int(hour), int(minute), int(second))

        # Find closest UUID by timestamp
        closest_uuid = None
        min_diff = None

        for uuid in uuid_list:
            try:
                # UUID format: TRI+52ca9b6a+2023-11-06-11h-33m-00s
                ts_part = uuid.split('+')[-1]
                uuid_ts = datetime.datetime.strptime(ts_part, '%Y-%m-%d-%Hh-%Mm-%Ss')
                time_diff = abs((episode_ts - uuid_ts).total_seconds())

                if min_diff is None or time_diff < min_diff:
                    min_diff = time_diff
                    closest_uuid = uuid
            except:
                continue

        return closest_uuid
    except:
        return None


def process_episode(episode, episode_idx, uuid, calib_loader, projector, cotracker, device,
                    max_frames=400, save_video=False, output_dir=None, use_batching=False, batch_size=300):
    """
    Process single DROID episode.
    Args:
        use_batching: If True, use batched tracking to reduce memory usage
        batch_size: Number of points per batch when use_batching=True (default 500)
    Returns:
        dict with processed data, or None if failed
    """
    # Get calibration (refined extrinsics + measured intrinsics)
    try:
        dual_params = calib_loader.get_dual_view_params(uuid, param_type='refined', require_refined=True)
        if dual_params is None:
            return None
    except:
        return None

    K_ext, E_ext = dual_params['exterior_1']
    K_wrist, E_wrist = dual_params['wrist']

    # First pass: count valid frames to check if episode is too long
    valid_frame_count = 0
    for step in episode['steps']:
        img_ext = step['observation']['exterior_image_1_left'].numpy()
        img_wrist = step['observation']['wrist_image_left'].numpy()
        if img_ext is not None and img_wrist is not None:
            if len(img_ext.shape) == 3 and len(img_wrist.shape) == 3:
                valid_frame_count += 1
                if valid_frame_count > max_frames:
                    # Episode is too long, skip it
                    return None

    # Check minimum length
    if valid_frame_count < 10:
        return None

    # Collect frames and observations
    frames_ext = []
    frames_wrist = []
    actions = []

    for step_idx, step in enumerate(episode['steps']):
        img_ext = step['observation']['exterior_image_1_left'].numpy()
        img_wrist = step['observation']['wrist_image_left'].numpy()

        if img_ext is None or img_wrist is None:
            continue
        if len(img_ext.shape) != 3 or len(img_wrist.shape) != 3:
            continue

        frames_ext.append(img_ext)
        frames_wrist.append(img_wrist)
        actions.append(step['action'].numpy())

    T = len(frames_ext)
    img_h, img_w = frames_ext[0].shape[:2]

    # Step 1: Compute mesh vertices for ALL frames (ground truth, not tracked)
    all_mesh_3d = []
    all_mesh_2d_ext = []
    all_mesh_vis_ext = []

    for action in actions:
        # Transform gripper offsets
        gripper_3d = transform_gripper_offsets(action)

        # Project to exterior camera
        mesh_2d, mesh_vis = projector._project_3d_to_2d(
            gripper_3d, K_ext, E_ext, img_h=img_h, img_w=img_w
        )

        all_mesh_3d.append(gripper_3d)
        all_mesh_2d_ext.append(mesh_2d)
        all_mesh_vis_ext.append(mesh_vis)

    all_mesh_2d_ext = np.array(all_mesh_2d_ext)  # [T, 7, 2]
    all_mesh_vis_ext = np.array(all_mesh_vis_ext)  # [T, 7]

    # Check if at least some mesh points visible in first frame
    if np.sum(all_mesh_vis_ext[0]) < 2:
        return None

    # Compute FIXED mesh for wrist camera (moves with gripper state, NOT tracked)
    all_mesh_2d_wrist_fixed = []
    for action in actions:
        gripper_state = action[6]
        mesh_wrist_fixed = get_wrist_fixed_mesh_2d(gripper_state, img_h, img_w)
        all_mesh_2d_wrist_fixed.append(mesh_wrist_fixed)
    all_mesh_2d_wrist_fixed = np.array(all_mesh_2d_wrist_fixed)  # [T, 7, 2]
    # Wrist mesh is always visible (hardcoded 2D, always in frame)
    all_mesh_vis_wrist_fixed = np.ones((T, 7), dtype=bool)

    # Step 2: Sample CoTracker query points at frame 0
    # LIBERO approach: 7 mesh + 98 grid + 1000 variance-sampled random = 1105 total
    mesh_2d_0 = all_mesh_2d_ext[0]  # [7, 2] - all mesh points from frame 0

    # Sample grid points for background coverage (uniform grid, always kept)
    grid_points_ext = sample_double_grid(n=7, img_h=img_h, img_w=img_w)  # [98, 2]

    # Sample random points across entire image (filtered by variance > 10.0)
    if episode_idx is not None:
        np.random.seed(episode_idx)
    random_points_ext = np.random.rand(1000, 2) * [img_w, img_h]
    random_points_ext = random_points_ext.astype(np.float32)

    # Combine: first 7 = mesh, next 98 = grid, last 1000 = random (variance filtered)
    query_points_ext = np.vstack([mesh_2d_0, grid_points_ext, random_points_ext])  # [1105, 2]
    num_mesh_ext = 7
    num_grid_ext = 98

    # Wrist: 7 tracked mesh + 98 grid + 1000 random = 1105 total tracked points
    # Fixed mesh is saved separately (not part of tracked points)
    # Get gripper mesh for first frame (initial tracking queries)
    gripper_state_0 = actions[0][6]  # action[6] is gripper state
    mesh_2d_wrist_tracked = get_wrist_gripper_mesh_2d(gripper_state_0, img_h, img_w)  # [7, 2]

    # Fixed 5x5 wrist query grid (closest to crop edges with light padding).
    fixed_grid_wrist = sample_uniform_grid(
        n=5,
        img_h=img_h,
        img_w=img_w,
        pad_ratio=0.10,
        crop_left=92.0,
        crop_right=272.0,
        crop_top=0.0,
        crop_bottom=180.0,
        out_w=224.0,
        out_h=224.0,
    )  # [25, 2], uniform in resized-crop coordinates
    # Keep 5 rows but compress vertical span so the bottom row sits higher (above gripper).
    fixed_grid_wrist_xy = fixed_grid_wrist.reshape(5, 5, 2)
    y_top = fixed_grid_wrist_xy[0, 0, 1]
    y_bottom = fixed_grid_wrist_xy[3, 0, 1]
    fixed_grid_wrist_xy[:, :, 1] = np.linspace(y_top, y_bottom, 5, dtype=np.float32)[:, None]
    fixed_grid_wrist = fixed_grid_wrist_xy.reshape(-1, 2)
    # Keep total wrist grid count at 98 for compatibility with existing downstream assumptions.
    support_grid_wrist = sample_double_grid(n=7, img_h=img_h, img_w=img_w)[:73]  # [73, 2]
    grid_points_wrist = np.vstack([fixed_grid_wrist, support_grid_wrist]).astype(np.float32)  # [98, 2]
    fixed_grid_wrist_indices = np.arange(7, 7 + len(fixed_grid_wrist), dtype=np.int32)  # [25]

    # Sample random points across entire image (filtered by variance > 10.0)
    if episode_idx is not None:
        np.random.seed(episode_idx + 1000)  # Different seed for wrist
    random_points_wrist = np.random.rand(1000, 2) * [img_w, img_h]
    random_points_wrist = random_points_wrist.astype(np.float32)

    # Combine: first 7 = tracked mesh, next 98 = grid, last 1000 = random (variance filtered)
    query_points_wrist = np.vstack([mesh_2d_wrist_tracked, grid_points_wrist, random_points_wrist])  # [1105, 2]
    num_mesh_wrist = 7
    num_grid_wrist = 98

    # Step 3: Run CoTracker with enhanced tracking (temporal sampling + variance filtering)
    # Exterior view
    video_ext_np = np.array(frames_ext).transpose(0, 3, 1, 2)  # [T, 3, H, W]
    video_ext_tensor = torch.from_numpy(video_ext_np).float() / 255.0
    video_ext_tensor = video_ext_tensor.unsqueeze(0).to(device)  # [1, T, 3, H, W]

    # Create temporal queries (mesh from t=0, others from random frames)
    queries_ext = create_temporal_queries(query_points_ext, T, num_mesh=7)
    queries_ext_tensor = torch.from_numpy(queries_ext).float().unsqueeze(0).to(device)

    # Track with variance filtering + forward/backward tracking
    # Threshold 10.0 matches LIBERO preprocessing
    # Preserve mesh (0-6) + grid (7-104) points, filter scattered (105-999)
    preserve_indices_ext = list(range(num_mesh_ext + num_grid_ext))  # 0-104

    # Choose tracking function based on batching flag
    track_fn = track_with_variance_filter_batched if use_batching else track_with_variance_filter

    if use_batching:
        tracks_ext, vis_ext = track_fn(
            cotracker, video_ext_tensor, queries_ext_tensor, device,
            var_threshold=10.0, preserve_indices=preserve_indices_ext, batch_size=batch_size
        )
    else:
        tracks_ext, vis_ext = track_fn(
            cotracker, video_ext_tensor, queries_ext_tensor, device,
            var_threshold=10.0, preserve_indices=preserve_indices_ext
        )
    # tracks_ext: [T, 1000, 2], vis_ext: [T, 1000]

    # Wrist view
    video_wrist_np = np.array(frames_wrist).transpose(0, 3, 1, 2)
    video_wrist_tensor = torch.from_numpy(video_wrist_np).float() / 255.0
    video_wrist_tensor = video_wrist_tensor.unsqueeze(0).to(device)

    # Create temporal queries (mesh + fixed wrist grid from t=0, others from random frames)
    queries_wrist = create_temporal_queries(query_points_wrist, T, num_mesh=7)
    queries_wrist[7:7 + len(fixed_grid_wrist), 0] = 0
    queries_wrist_tensor = torch.from_numpy(queries_wrist).float().unsqueeze(0).to(device)

    # Track with variance filtering + forward/backward tracking (1105 points)
    # Threshold 10.0 matches LIBERO preprocessing
    # Preserve mesh (0-6) + grid (7-104) points, filter random (105-1104)
    preserve_indices_wrist = list(range(num_mesh_wrist + num_grid_wrist))  # 0-104

    if use_batching:
        tracks_wrist, vis_wrist = track_fn(
            cotracker, video_wrist_tensor, queries_wrist_tensor, device,
            var_threshold=10.0, preserve_indices=preserve_indices_wrist, batch_size=batch_size
        )
    else:
        tracks_wrist, vis_wrist = track_fn(
            cotracker, video_wrist_tensor, queries_wrist_tensor, device,
            var_threshold=10.0, preserve_indices=preserve_indices_wrist
        )
    # tracks_wrist: [T, 1105, 2], vis_wrist: [T, 1105]
    # First 7 = tracked mesh, next 98 = grid, rest = variance-filtered random
    # Fixed mesh is saved separately in all_mesh_2d_wrist_fixed

    # Step 4: Save preview video if requested
    if save_video and output_dir:
        video_frames = []
        for t in range(T):
            # Exterior view
            viz_ext = frames_ext[t].copy()

            # Draw ground truth mesh vertices (blue circles)
            for i in range(7):
                if all_mesh_vis_ext[t, i]:
                    pt = tuple(all_mesh_2d_ext[t, i].astype(int))
                    cv2.circle(viz_ext, pt, 5, (255, 0, 0), 2)  # Blue hollow
                    cv2.putText(viz_ext, str(i), (pt[0]+6, pt[1]-6),
                               cv2.FONT_HERSHEY_SIMPLEX, 0.3, (255, 0, 0), 1)

            # Draw tracked mesh vertices (red filled) - first 7 of CoTracker
            for i in range(7):
                if vis_ext[t, i]:
                    pt = tuple(tracks_ext[t, i].astype(int))
                    cv2.circle(viz_ext, pt, 3, (0, 0, 255), -1)  # Red filled

            # Draw other CoTracker points (green, smaller)
            for i in range(7, len(tracks_ext[t])):
                if vis_ext[t, i]:
                    pt = tuple(tracks_ext[t, i].astype(int))
                    cv2.circle(viz_ext, pt, 1, (0, 255, 0), -1)

            cv2.putText(viz_ext, f"Ext: GT mesh (blue) | Tracked mesh (red) | Others (green)",
                       (5, 15), cv2.FONT_HERSHEY_SIMPLEX, 0.35, (255, 255, 255), 1)

            # Wrist view
            viz_wrist = frames_wrist[t].copy()

            # Draw TRACKED mesh vertices (first 7 points) - magenta filled circles
            for i in range(7):
                if vis_wrist[t, i]:
                    pt = tuple(tracks_wrist[t, i].astype(int))
                    cv2.circle(viz_wrist, pt, 3, (255, 0, 255), -1)  # Magenta filled

            # Draw FIXED mesh vertices (from separate array) - cyan hollow circles
            for i in range(7):
                if all_mesh_vis_wrist_fixed[t, i]:
                    pt = tuple(all_mesh_2d_wrist_fixed[t, i].astype(int))
                    cv2.circle(viz_wrist, pt, 4, (255, 255, 0), 2)  # Cyan hollow
                    # Label finger tips and joints
                    if i in [1, 2, 3, 4]:  # Tips and joints
                        cv2.putText(viz_wrist, str(i), (pt[0]+5, pt[1]-5),
                                   cv2.FONT_HERSHEY_SIMPLEX, 0.25, (255, 255, 0), 1)

            # Draw other CoTracker points (grid + random, green, smaller)
            for i in range(7, len(tracks_wrist[t])):
                if vis_wrist[t, i]:
                    pt = tuple(tracks_wrist[t, i].astype(int))
                    cv2.circle(viz_wrist, pt, 1, (0, 255, 0), -1)  # Green

            cv2.putText(viz_wrist, f"Wrist: Tracked mesh (magenta) | Fixed mesh (cyan) | Others (green)",
                       (5, 15), cv2.FONT_HERSHEY_SIMPLEX, 0.3, (255, 255, 255), 1)

            # Concatenate side by side
            combined = np.concatenate([viz_ext, viz_wrist], axis=1)
            cv2.putText(combined, f"Episode {episode_idx} | Frame {t}/{T}",
                       (combined.shape[1]//2 - 60, 15),
                       cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 255, 0), 2)

            video_frames.append(combined)

        video_path = output_dir / f"preview_episode_{episode_idx:06d}.mp4"
        media.write_video(str(video_path), video_frames, fps=10)

    # Return processed data
    return {
        'episode_idx': episode_idx,
        'uuid': uuid,
        'frames_exterior': np.array(frames_ext),  # [T, H, W, 3]
        'frames_wrist': np.array(frames_wrist),
        'actions': np.array(actions),  # [T, 7]
        'mesh_vertices_2d_exterior': all_mesh_2d_ext,  # [T, 7, 2] - ground truth (3D projection)
        'mesh_vertices_vis_exterior': all_mesh_vis_ext,  # [T, 7]
        'mesh_vertices_2d_wrist_fixed': all_mesh_2d_wrist_fixed,  # [T, 7, 2] - fixed mesh (2D, hardcoded from annotations)
        'mesh_vertices_vis_wrist_fixed': all_mesh_vis_wrist_fixed,  # [T, 7]
        'tracks_exterior': tracks_ext,  # [T, 1105, 2] - 7 mesh + 98 grid + ~1000 variance-filtered random
        'tracks_vis_exterior': vis_ext,  # [T, 1105]
        'tracks_wrist': tracks_wrist,  # [T, 1105, 2] - 7 tracked mesh + 98 grid + ~1000 variance-filtered random
        'tracks_vis_wrist': vis_wrist,  # [T, 1105] - Fixed mesh saved separately above
        'wrist_fixed_grid_points': fixed_grid_wrist,  # [25, 2] in original 180x320 frame space
        'wrist_fixed_grid_indices': fixed_grid_wrist_indices,  # [25] indices into tracks_wrist axis=1
    }


def main():
    import argparse
    parser = argparse.ArgumentParser()
    parser.add_argument('--num-episodes', type=int, default=None, help='Number of episodes to process (deprecated, use --end-episode)')
    parser.add_argument('--start-episode', type=int, default=0, help='Start episode index')
    parser.add_argument('--end-episode', type=int, default=None, help='End episode index (exclusive)')
    parser.add_argument('--episode-ids-file', type=str, default=None, help='JSON file with specific episode IDs to process')
    parser.add_argument('--output-dir', type=str, default='/tmp/droid_rlds_final', help='Output directory')
    parser.add_argument('--max-frames', type=int, default=400, help='Max frames per episode')
    parser.add_argument('--save-previews', type=int, default=3, help='Number of preview videos to save')
    parser.add_argument('--gpu-id', type=int, default=None, help='GPU ID for logging')
    args = parser.parse_args()

    # Handle backward compatibility
    if args.end_episode is None and args.num_episodes is not None:
        args.end_episode = args.start_episode + args.num_episodes

    # Load episode IDs from file if provided
    episode_ids_list = None
    if args.episode_ids_file:
        import json
        with open(args.episode_ids_file, 'r') as f:
            episode_ids_list = json.load(f)
        print(f"Loaded {len(episode_ids_list)} episode IDs from {args.episode_ids_file}")

    output_dir = Path(args.output_dir)
    output_dir.mkdir(parents=True, exist_ok=True)
    preview_dir = output_dir / 'preview_videos'
    preview_dir.mkdir(exist_ok=True)
    data_dir = output_dir / 'data'
    data_dir.mkdir(exist_ok=True)

    gpu_label = f"GPU {args.gpu_id}" if args.gpu_id is not None else "Processing"

    print("=" * 80)
    print(f"DROID Preprocessing: {gpu_label}")
    print("=" * 80)
    if episode_ids_list:
        print(f"  Mode: Direct episode access (from file)")
        print(f"  Episodes to process: {len(episode_ids_list)}")
    else:
        print(f"  Mode: Range iteration")
        print(f"  Episode range: {args.start_episode} to {args.end_episode if args.end_episode else 'end'}")
    print(f"  Max frames: {args.max_frames}")
    print(f"  Output: {output_dir}")
    print(f"  Preview videos: {args.save_previews}")
    print("=" * 80)

    # Initialize
    # CUDA_VISIBLE_DEVICES handles GPU selection, so always use cuda:0
    # (gpu_id is only for logging purposes)
    device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')

    calib_dir = '/root/workspace/code/wmrl/Dual-Dynamics-Models/DROID-main/vision/u/wenlongh/datasets/droid_v4/cameras'
    calib_loader = CameraCalibrationLoader(calib_dir)
    projector = FrankaMeshProjector(use_gui=False)
    cotracker, device = load_cotracker(device=device)

    calib_path = Path(calib_dir)
    uuid_list = [f.stem.replace('_cameras', '') for f in sorted(calib_path.glob("*_cameras.json"))]
    print(f"Loaded {len(uuid_list)} camera calibrations")

    # Load dataset
    droid_path = '/mnt/kevin/data/droid/droid/1.0.0'
    print("Loading DROID dataset...")
    builder = tfds.builder_from_directory(droid_path)
    dataset = builder.as_dataset(split='train')

    # Process episodes
    processed_count = 0
    skipped_count = 0

    # DIRECT EPISODE ACCESS MODE (from file)
    if episode_ids_list:
        print(f"\n{gpu_label}: Processing {len(episode_ids_list)} specific episodes...")
        pbar = tqdm(total=len(episode_ids_list), desc=gpu_label)

        for episode_idx in episode_ids_list:
            pbar.update(1)

            # Skip if already processed
            npz_path = data_dir / f"episode_{episode_idx:06d}.npz"
            if npz_path.exists():
                continue

            # Skip directly to this episode
            episode = dataset.skip(episode_idx).take(1)
            episode = next(iter(episode))

            # Find calibration
            uuid = find_closest_calibration(episode, uuid_list)
            if uuid is None or not calib_loader.has_refined_extrinsics(uuid):
                skipped_count += 1
                continue

            # Process episode (no need to check length - already filtered!)
            save_video = processed_count < args.save_previews

            try:
                # OOM retry logic: try unbatched first, retry with batching on OOM
                try:
                    result = process_episode(
                        episode, episode_idx, uuid, calib_loader, projector,
                        cotracker, device, max_frames=args.max_frames,
                        save_video=save_video, output_dir=preview_dir,
                        use_batching=False
                    )
                except torch.cuda.OutOfMemoryError:
                    print(f"\n  OOM on episode {episode_idx}, retrying with batching (batch_size=600)...")
                    torch.cuda.empty_cache()
                    result = process_episode(
                        episode, episode_idx, uuid, calib_loader, projector,
                        cotracker, device, max_frames=args.max_frames,
                        save_video=save_video, output_dir=preview_dir,
                        use_batching=True, batch_size=150
                    )

                if result is None:
                    skipped_count += 1
                    continue

                # Save as NPZ
                npz_path = data_dir / f"episode_{episode_idx:06d}.npz"
                np.savez_compressed(
                    npz_path,
                    episode_idx=result['episode_idx'],
                    uuid=result['uuid'],
                    images_exterior=result['frames_exterior'],
                    images_wrist=result['frames_wrist'],
                    actions=result['actions'],
                    mesh_vertices_2d_exterior=result['mesh_vertices_2d_exterior'],
                    mesh_vertices_vis_exterior=result['mesh_vertices_vis_exterior'],
                    mesh_vertices_2d_wrist_fixed=result['mesh_vertices_2d_wrist_fixed'],
                    mesh_vertices_vis_wrist_fixed=result['mesh_vertices_vis_wrist_fixed'],
                    tracks_exterior=result['tracks_exterior'],
                    tracks_vis_exterior=result['tracks_vis_exterior'],
                    tracks_wrist=result['tracks_wrist'],
                    tracks_vis_wrist=result['tracks_vis_wrist'],
                    wrist_fixed_grid_points=result['wrist_fixed_grid_points'],
                    wrist_fixed_grid_indices=result['wrist_fixed_grid_indices'],
                    mesh_indices=np.array([0, 1, 2, 3, 4, 5, 6], dtype=np.int32),
                )

                processed_count += 1

            except Exception as e:
                print(f"\nError processing episode {episode_idx}: {e}")
                skipped_count += 1
                continue

        pbar.close()

    # RANGE ITERATION MODE (backward compatibility)
    else:
        # EFFICIENT SKIPPING: Use dataset.skip() instead of manual iteration
        if args.start_episode > 0:
            print(f"{gpu_label} Skipping to episode {args.start_episode}...")
            dataset = dataset.skip(args.start_episode)

        if args.end_episode is not None:
            num_episodes = args.end_episode - args.start_episode
            print(f"{gpu_label} Taking {num_episodes} episodes...")
            dataset = dataset.take(num_episodes)

        total_to_process = (args.end_episode - args.start_episode) if args.end_episode else None
        pbar = tqdm(total=total_to_process, desc=gpu_label)

        for local_idx, episode in enumerate(dataset):
            # Calculate actual episode index in full dataset
            episode_idx = args.start_episode + local_idx

            # Skip if already processed
            npz_path = data_dir / f"episode_{episode_idx:06d}.npz"
            if npz_path.exists():
                pbar.update(1)
                continue

            # Find calibration
            uuid = find_closest_calibration(episode, uuid_list)
            if uuid is None or not calib_loader.has_refined_extrinsics(uuid):
                skipped_count += 1
                pbar.update(1)
                continue

            # Check episode length
            episode_length = sum(1 for _ in episode['steps'])
            if episode_length > args.max_frames or episode_length < 10:
                skipped_count += 1
                pbar.update(1)
                continue

            # Process episode
            save_video = processed_count < args.save_previews

            try:
                # OOM retry logic: try unbatched first, retry with batching on OOM
                try:
                    result = process_episode(
                        episode, episode_idx, uuid, calib_loader, projector,
                        cotracker, device, max_frames=args.max_frames,
                        save_video=save_video, output_dir=preview_dir,
                        use_batching=False
                    )
                except torch.cuda.OutOfMemoryError:
                    print(f"\n  OOM on episode {episode_idx}, retrying with batching (batch_size=150)...")
                    torch.cuda.empty_cache()
                    result = process_episode(
                        episode, episode_idx, uuid, calib_loader, projector,
                        cotracker, device, max_frames=args.max_frames,
                        save_video=save_video, output_dir=preview_dir,
                        use_batching=True, batch_size=150
                    )

                if result is None:
                    skipped_count += 1
                    pbar.update(1)
                    continue

                # Save as NPZ
                npz_path = data_dir / f"episode_{episode_idx:06d}.npz"
                np.savez_compressed(
                    npz_path,
                    episode_idx=result['episode_idx'],
                    uuid=result['uuid'],
                    images_exterior=result['frames_exterior'],
                    images_wrist=result['frames_wrist'],
                    actions=result['actions'],
                    mesh_vertices_2d_exterior=result['mesh_vertices_2d_exterior'],
                    mesh_vertices_vis_exterior=result['mesh_vertices_vis_exterior'],
                    mesh_vertices_2d_wrist_fixed=result['mesh_vertices_2d_wrist_fixed'],
                    mesh_vertices_vis_wrist_fixed=result['mesh_vertices_vis_wrist_fixed'],
                    tracks_exterior=result['tracks_exterior'],
                    tracks_vis_exterior=result['tracks_vis_exterior'],
                    tracks_wrist=result['tracks_wrist'],
                    tracks_vis_wrist=result['tracks_vis_wrist'],
                    wrist_fixed_grid_points=result['wrist_fixed_grid_points'],
                    wrist_fixed_grid_indices=result['wrist_fixed_grid_indices'],
                    mesh_indices=np.array([0, 1, 2, 3, 4, 5, 6], dtype=np.int32),
                )

                processed_count += 1
                pbar.update(1)

            except Exception as e:
                print(f"\nError processing episode {episode_idx}: {e}")
                skipped_count += 1
                pbar.update(1)
                continue

        pbar.close()

    # Save metadata
    metadata = {
        'num_episodes': processed_count,
        'split': 'train',
        'camera_params': {
            'exterior': {
                'extrinsics': 'refined',
                'intrinsics': 'measured',
                'inversion': False
            },
            'wrist': {
                'sampling': 'random',
                'dense_region': 'bottom_60_100_pct'
            }
        },
        'point_distribution': {
            'exterior': {
                'total_tracked_points': 1000,
                'mesh_vertices_tracked': 7,  # First 7 indices
                'additional_points': 993,  # Indices 7-999
                'arm_shaped_strategy': 'gaussian_15px_per_mesh + lines_between',
                'note': 'Mesh vertices tracked by CoTracker AND saved as ground truth separately'
            },
            'wrist': {
                'total_tracked_points': 1000,
                'sparse_uniform': 300,  # First 300 indices
                'dense_bottom': 700  # Last 700 indices in bottom 60%-100%
            }
        },
        'image_resolution': [180, 320],
        'max_frames_per_episode': args.max_frames,
        'cotracker_model': 'scaled_offline.pth',
        'cotracker_chunking': 'automatic_internal_only'
    }

    with open(output_dir / 'metadata.json', 'w') as f:
        json.dump(metadata, f, indent=2)

    print("\n" + "=" * 80)
    print("Preprocessing Complete")
    print("=" * 80)
    print(f"  Processed: {processed_count} episodes")
    print(f"  Skipped: {skipped_count} episodes")
    print(f"  Preview videos: {preview_dir}")
    print(f"  NPZ data: {data_dir}")
    print(f"  Metadata: {output_dir / 'metadata.json'}")
    print("=" * 80)


if __name__ == "__main__":
    main()