[Question] Need some clarity on reference frames usage in ArticulationData

[kbkartik]

Question

newbie to sim2real world. I am trying to understand how reference frames are setup in ArticulationData. My doubt is w.r.t the following 6 methods:

1. root_link_state_w()
2. root_com_state_w()
3. body_link_state_w()
4. body_com_state_w()
5. com_pos_b()
6. com_quat_b()
7. root_com_lin_vel_b()
8. root_com_ang_vel_b()

Among all these methods, I am seeing multiple definitions of reference frame such as:

1. simulation world frame
2. body frame
3. center of mass frame
4. actor frame
5. orientation of principle inertia (or principle axes of inertia)

I know what reference frames are. What I don't understand is how its setup in IsaacLab? For instance,

1. Why do we adjust linear velocity in root_link_state_w() and (pos, quat) in root_com_state_w()?
2. Why do we use com_pos_b and com_quat_b in root_link_state_w() and root_com_state_w()? Are we transforming from body frame to world frame?
3. Why do we do update_articulations_kinematics() in body_link_state_w() and body_com_state_w()?
4. What is the mapping/ordering of bodies in body_link_state_w/body_com_state_w's shape: (num_instances, num_bodies, 13)?
5. In root_com_lin_vel_b() and root_com_ang_vel_b(), why are we using root_link_quat_w to inverse rotate root_com_lin_vel_b and root_com_ang_vel_b? Shouldn't we be using root_com_quat_w? Is this a bug?

Some clarity on the above questions will definitely be helpful to understand the code and the documentations.

best,
kb

[RandomOakForest]

Thank you for submitting these questions. We are working on improving our documentation. I will move this to a discussion for the team to follow up.

[StrainFlow]

Great questions, I've had similar trouble with the clarity of these functions. I'm going to try clarifying these in the documentation. Would you mind if I shared links to the revisions as I go and have you proof read them for me for clarity? Once they're all clear we can submit a PR. This is an open discussion as well, so I would welcome feedback from anyone else as we go.

[kbkartik]

Works with me!

[kbkartik]

@StrainFlow rechecking with you on this. Thanks!

[RandomOakForest]

In a note to give some context, the physics engine provides us pose data in link frame and velocity data in com frame. Anything that has link in the name is in link frame. For velocities, we do the transformation under the hood to convert from com frame to link frame. Similar idea for APIs that have com in the name.

[kbkartik]

I understand the physical meaning behind com. From my understanding, given a rigid body:

1. frame and space both mean the same thing.
2. We take the com of the rigid body and assign it to the origin. This becomes my body frame B.
3. Now, given a rotation matrix R_{BW} (world frame w.r.t body frame) and a translation vector, I can transform my rigid body to the world frame.

If my understanding is correct according to you, I have the following questions:

1. Is this how com frame defined?
2. What is the physical meaning behind a link frame?
3. Why does the physics engine provide pose data in link frame and velocity data in com frame? Why not both in the same frame? Is there an underlying specific reason?
4. How are link and com related to root and body?

[StrainFlow]

You can transform between the com frame and the world frame with just a rotation matrix, no translation.

The link properties are built on the same coordinate frame as the parent joint for a given link in the articulation.

COM and Link refer coordinate frames while root and body refer to rigid bodies. The root always gives you the properties of the articulation root where the body properties give you a dictionary you can use to query the data from any rigid body in the articulation.

[StrainFlow]

There's not a specific reason for why the physics engine provides the data the way it does, it was just never an issue until recently with robotics.

[kbkartik]

The link properties are built on the same coordinate frame as the parent joint for a given link in the articulation.

So, let's assume a 3 link planar arm attached to a base link, say at (1,1). Also, assume the world coordinate frame to be the origin.

1. The root becomes the base link, with a reference frame w.r.t world, say origin.

2. Link 3 will use the parent joint's (the joint connecting Link 3 and Link 2) reference frame which is defined w.r.t world.

3. Given a vector defined in the world frame, I can define the same vector in the Link 3's frame by multiplying the inverse of the rotation matrix defined in (2).

Am I correct in my understanding above?

Thanks!