Some Doubt about "action" and "state"
Hello, sorry to bother you. I have some questions I'd like to ask you.
In the info.json, there is:
"axes": [
"x",
"y",
"z",
"rot6d_0",
"rot6d_1",
"rot6d_2",
"rot6d_3",
"rot6d_4",
"rot6d_5",
"gripper"
]
- If I want to use the "action" to step the libero env, how can i set the controller and transform the "action"?
- What is the coordinate frames of "action" and "observation.state"? robot0_base/world/camera？
- lerobot_libero_mem_abs, what's the meaning of 'abs'?
I created this dataset specifically for use with XVLA in LeRobot. Since XVLA is pretrained on ee6d actions, I used the end-effector data from the HDF5 files so that the coordinate frames should match those used in the LIBERO mem dataset.
I’m not entirely sure what you mean by “setting the controller” or “transforming the action.” I am using LIBERO mem through LeRobot, so I did not implement a separate controller or apply any additional action transformations myself.
Unfortunately, I haven’t been able to successfully train XVLA yet, so I can’t fully confirm whether the dataset conversion is correct. If you manage to obtain results using this dataset, I’d really appreciate it if you could let me know how you did it.
Thank you for your answer. I have clarified the above issues.
Furthermore, I have another question regarding the last dimension of the "observation.state" item in this dataset—it appears to be problematic.
Does it represent the gripper's qpos(normalized and represented in -1~1, which -1 for open and 1 for close)? It seems to assign the control variable of the action to the gripper.
By playing back the recorded trajectories and comparing the states in the dataset with the environmental observations, it was discovered that the gripper dimension of the states seems to have an issue.（already confirm the pose dimensions of the dataset state is correct）