Questions about the statistical data in demo/stat.json file

[miznchimaki]

Excellent work! Thank you for sharing the code.

While reading the source code for the inference demo, I noticed that the action inputs are normalized using statistical values (e.g., min/max or p01/p99). I have a few questions regarding the statistics provided in demo/stat.json:

1. How were the min/max, p01/p99, and mean/std values for the robot arm joints and end-effector poses obtained? Were these statistics computed from the training data?

2. What exactly does action represent in action_joint and action_pose? Does it refer to delta actions computed as the difference between the robot state readings at two consecutive frames, i.e., (s_{t+1} - s_t)?

3. For state_pose and action_pose, excluding the gripper dimension, do the remaining left+right 6-dimensional representation are [x, y, z] + Euler angles? If so, are the Euler angles represented in roll-pitch-yaw (RPY) order?

Thank you in advance!

[miznchimaki]

For state_pose and action_pose, is the 14-dimensional vector left-[x, y, z, wx, wy, wz, gripper] + right-[x, y, z, wx, wy, wz, gripper], which is written in the LeRobot documentation?
https://huggingface.co/docs/lerobot/action_representations