Firstly, I would like to thank you for your outstanding project and the well-documented code!
I have a query regarding the provided example for a single robot arm. I noticed that the key robot0_eef_rot_axis_angle_wrt_start in the shape meta represents the robot's end effector orientation relative to the episode's starting orientation. From the paper, it seems that you have used relative transforms. Should this key be removed when a relative trajectory is used?
Additionally, I observed in the code that when relative transforms are used, the last position obs_dict[f"robot{robot_id}_eef_pos"][-1] of the end effector in the history before normalization appears to be always set to an identity matrix. Is this an intentional design for compatibility with other representations, or could it be a bug?
Good afternoon,
Firstly, I would like to thank you for your outstanding project and the well-documented code!
I have a query regarding the provided example for a single robot arm. I noticed that the key robot0_eef_rot_axis_angle_wrt_start in the shape meta represents the robot's end effector orientation relative to the episode's starting orientation. From the paper, it seems that you have used relative transforms. Should this key be removed when a relative trajectory is used?
Additionally, I observed in the code that when relative transforms are used, the last position obs_dict[f"robot{robot_id}_eef_pos"][-1] of the end effector in the history before normalization appears to be always set to an identity matrix. Is this an intentional design for compatibility with other representations, or could it be a bug?