Even though it varies with robot configuration, what exactly do the robot state and action entail? For example, in OpenVLA they are very clear that the output robot action is represented as delta-end effector in cartesian coordinates, i.e., deltas of [x, y, z, rotation_x, rotation_y, rotation_z, gripper]
What does Pi output? And how are the states represented? Are they represented as [j1, j2, j3, j4, j5, j6, x, y, z, qx, qy, qz, qw gripper]?
In openVLA the deltas are in robot base frame and not in gripper frame. What is the case for PiZero? Any help is appreciated.
Reference for OpenVLA action with timestamp: https://youtu.be/-0s0v3q7mBk?t=675