State/Action representation in LIBERO dataset (EEF vs Joint)

[lyfeng001]

Hello, thanks for your great work first!

I recently downloaded the LIBERO dataset, and I noticed something confusing regarding the representation of state and action in the feature files.

• The documentation suggests that both state and action are represented in EEF (end-effector) space.
• Since Panda is a 7-DOF arm, it makes sense that state could be represented as a 6D pose (position + orientation) plus 2D gripper values.
• However, when I inspected the data, I found that the state values are numerically very far from the corresponding actions. Intuitively, since actions are control inputs and states are feedback signals, their values should be in a similar range.

• I also noticed that some dimensions of state are very close to 3.14, which makes me suspect that the state vector might actually be in joint space.

Additionally:

• In Issue #95, it was mentioned that the state is in joint representation.
• But if this is the case, then applying delta-action would essentially compute EEF - Joint, which does not have a clear physical meaning.
• On the other hand, during inference the model input is constructed as:

np.concatenate(
    (
        obs["robot0_eef_pos"],
        _quat2axisangle(obs["robot0_eef_quat"]),
        obs["robot0_gripper_qpos"],
    )
)

This suggests that training should also be using EEF representation.

These points seem contradictory. Could you please clarify:

1. What is the exact representation used for state and action in the dataset?
2. Does delta-action have a meaningful interpretation in the current setup?
3. What representation should be assumed during inference and training?

Thanks a lot for your help!