Real-world datasets, engineered at AI scale.

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

class BlueProjectsEgoDataset(Dataset):
    def __init__(self, hdf5_path):
        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 IMU
        self.mano_hands = self.file['annotations/mano_3d_joints']

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

# Stream 100k Hours Egocentric POV Vault Sample
dataset = BlueProjectsEgoDataset('s3://blueprojects-vault/egocam_sample.hdf5')
print(f"VLA Training Ready: {len(dataset)} episodes")

{
  "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],
  "anonymization": { "faces_blurred": true, "pii_scrubbed": true }
}

/
├── observations
│   ├── images (cam_high 1920x1080 @ 60fps RGB)
│   └── qpos (14-DOF Leader Joint State Angles)
├── action (14-DOF Follower Motor Target Positions)
└── telemetry (6-Axis Force Torque Arrays + Microsecond IMU)