import torch
from torch.utils.data import Dataset
import h5py

class BlueProjectsEgoDataset(Dataset):
    def __init__(self, hdf5_path, transform=None):
        self.file = h5py.File(hdf5_path, 'r')
        self.cam_rgb = self.file['observations/images/cam_high']
        self.imu_6dof = self.file['telemetry/imu']  # Microsecond synced 6-DOF
        self.mano_hands = self.file['annotations/mano_3d_joints']  # 21 Keypoints

    def __getitem__(self, idx):
        rgb_frame = torch.from_numpy(self.cam_rgb[idx])
        imu_sample = torch.from_numpy(self.imu_6dof[idx])
        mano_pose = torch.from_numpy(self.mano_hands[idx])
        return {'rgb': rgb_frame, 'imu': imu_sample, 'mano_3d': mano_pose}

# Load 100,000+ Hours Egocentric POV Vault Sample
dataset = BlueProjectsEgoDataset('s3://blueprojects-vault/egocam_sample_01.hdf5')
print(f"Dataset Frames: {len(dataset)} | Ready for VLA & Policy Training")

{
  "t_us": 907814846,
  "device_id": "EgoCAM-6DOF-0482",
  "acc": [0.201113, -0.483629, -10.649409],
  "gyro": [-0.116047, 0.026616, 0.072396],
  "gaze_vector_200hz": [0.421, 0.819, 0.104],
  "mano_3d_hand_pose": {
    "left_hand": [[0.01, 0.02, 0.03], "... 21 Joint Keypoints ..."],
    "right_hand": [[0.04, 0.05, 0.06], "... 21 Joint Keypoints ..."]
  },
  "anonymization": { "faces_blurred": true, "pii_scrubbed": true }
}

# 50Hz Bilateral ALOHA HDF5 Teleoperation Hierarchy Spec
/
├── observations
│   ├── images
│   │   ├── cam_high (1920x1080 @ 60fps RGB)
│   │   ├── cam_left_wrist (1280x720 RGB)
│   │   └── cam_right_wrist (1280x720 RGB)
│   └── qpos (14-DOF Leader Joint State Angles)
├── action (14-DOF Follower Motor Target Positions)
├── telemetry
│   ├── load_cell_effort (6-Axis Force Torque Arrays)
│   └── imu (Microsecond Timestamps)
└── metadata (Consented Operative ID, Factory ID, Trade Cluster)