Action space clarification

[abhyudit309]

Hi,

I wanted to confirm that the action space consists of the end-effector position and orientation displacement (along with the gripper state). Are these displacements/delta values in the world frame or the end-effector frame? Also, is there some way to know what the initial position of the robot arm is (from the datasets)?

Thanks!

[peter-mitrano-bg]

Did you ever figure this out? I think knowing this is very important, and the original octo paper does not state the coordinate frame clearly.

[JemuelStanley47]

Please correct me if you have already figured this out. According to some of the datasets in Open X Embodiment Dataset, the action space is w.r.t the robot base frame.

For example:
Berkeley Autolab UR5 (https://sites.google.com/view/berkeley-ur5/home ) specifies that the action:
x,y,z, qx,qy,qz,qw is the end-effector pose expressed in the robot base frame

[peter-mitrano-bg]

I have looked at some of the datasets comprising OXE and I'm not convinced they are all the same... but yes I agree that most of them are deltas in the "base frame" (which will be different for every robot of course)