"action": state_qpos.reshape((epi_len, -1)),

[RuiboLiu77]

Hello, thank you very much for your open-source project.
I got a question about the hdf5 data loading process, I notice that in function in below:

def load_hdf5_dataset(
episode_path: str | Path,
) -> dict:
"""Load hdf5 dataset and return a dict with observations and actions"""

with h5py.File(episode_path) as f:
    state_images_cam_high = np.array(f["observations/images/cam_high"])
    state_images_cam_left_wrist = np.array(f["observations/images/cam_left_wrist"])
    state_images_cam_right_wrist = np.array(f["observations/images/cam_right_wrist"])
    state_qpos = np.array(f["observations/qpos"])

assert (
    state_images_cam_high.shape[0]
    == state_images_cam_left_wrist.shape[0]
    == state_images_cam_right_wrist.shape[0]
    == state_qpos.shape[0]
)

epi_len = state_images_cam_high.shape[0]
episode = {
    "observation.state": state_qpos.reshape((epi_len, -1)),
    "observation.images.top_head": state_images_cam_high,
    "observation.images.hand_left": state_images_cam_left_wrist,
    "observation.images.hand_right": state_images_cam_right_wrist,
    "action": state_qpos.reshape((epi_len, -1)),
    "epi_len": epi_len
}
return episode

you use state_qpos as action input, is that a special design or a mini bug? That cofuse me a bit
Thank you for your attention.

[Tradewindycc]

Hi, thanks for your question. Yes, that's a co-design for both data collection and model training .
During data collection stage, we save the slave joint angles as 'action' in HDF5, and the master joint angles as 'observation/qpos'. For ALOHA setup, the master joint angles('observation/qpos') are actually the action commands for slave joints, and we expect the model to predict action command rather than proprioceptions (which is 'action' in HDF5 in practice).