Some questions regarding force/torque for actor and link

[nv-quan]

First, thank you for these amazing projects (both SAPIEN and ManiSkill). I have several questions while building my own simulated environment:

• In ManiSkill, cuda_rigid_body_force is used to apply a force to an actor. I am wondering what the difference is between cuda_rigid_body_force and cuda_rigid_dynamic_force and what are the meanings of these arrays?
• Is it possible to apply a force and/or a torque to a link inside an articulation by setting cuda_rigid_body_force and cuda_rigid_body_torque similar to the ManiSkill implementation. Would this override the force/torque values the link is already experiencing (due to other links, gravity, etc.)?
• Is it possible to query the position of a contact force in GPU simulation, similar to contact.points in CPU simulation? I couldn't find how to do it.
  Thanks a lot!

[StoneT2000]

• should probably detail this in our documentation. I took a look at the SAPIEN code to figure it out. Essentially all things cuda_rigid_dynamic_x are slices of the cuda_rigid_body_x data. Specifically rigid_body contains the 13D pose+velocities of all rigid body actors, and rigid body links in articulations (and further padded with zeros based on the articulation with the largest DOF). rigid_dynamic only refers to actors.

SAPIEN/src/physx/physx_system.cpp

Lines 1034 to 1047 in fd97712

	void PhysxSystemGpu::allocateCudaBuffers() {
	SAPIEN_PROFILE_FUNCTION;
	int rigidDynamicCount = mRigidDynamicComponents.size();
	int rigidBodyCount = rigidDynamicCount + mGpuArticulationCount * mGpuArticulationMaxLinkCount;
	ensureCudaDevice();
	// rigid body data buffer
	mCudaRigidBodyBuffer = CudaArray({rigidBodyCount, 13}, "f4");
	mCudaRigidDynamicHandle = CudaArrayHandle{.shape = {rigidDynamicCount, 13},
	.strides = {52, 4},
	.type = "f4",
	.cudaId = mCudaRigidBodyBuffer.cudaId,
	.ptr = (float *)mCudaRigidBodyBuffer.ptr};

• There is gpuApplyRigidDynamic Data/Force/Torque available in the SAPIEN API. For actors we have exposed the data (pose/velocities) and force apis, but not Torque (you can still use it but a nice more readable api without having to deal with the cuda_x buffers is not available at the moment). As far as I know you can't do this for links.
• In GPU sim you can't query contact points at the moment. If you really need it you can consider running in GPU sim first, then re-run the same actions in the CPU sim

[nv-quan]

Thanks a lot for the detailed answer and sorry for my late response! If you don't mind, I'd like to have a few follow-ups:

• Is there a way to get the net force applied on an actor? I want to query the net force, or at least, the total force that one actor is acting on another. I tried querying the contact force with gpu_query_contact_pair_impulses but it doesn't seem to include other forces in the interaction (e.g. friction).
• I think apply_torque should be very similar to apply_force, but please let me know if my implementation is correct (for GPU only):

def apply_torque(actor, torque: Array):
    actor.px.cuda_rigid_body_torque.torch()[
        actor._body_data_index, :3
    ] = common.to_tensor(torque, device=actor.device)
    actor.px.gpu_apply_rigid_dynamic_torque()

• As you said, currently there is no way to apply a perturbation force/torque to a link inside a robot, but is there a plan to support this in the future? I believe this could be useful in various of scenarios.

[StoneT2000]

1. For different kind of forces see https://maniskill.readthedocs.io/en/latest/user_guide/tutorials/custom_tasks/advanced.html#contact-forces-on-the-gpu-cpu, there is a net force API. I'm not sure what the exact code is used by physx to generate that though
2. That implementation looks right to me, you can give it a try
3. This is dependent on the physics backend's capabilities. The ManiSkill team is not directly working on adding new features to the rigid-body physics engine at this point in time

[nv-quan]

Thank you! For 1, the actor.get_net_contact_forces() api also doesn't include friction. I ended up using F = ma to calculate the net force where acceleration a is estimated by calculating the velocity difference before and after simulation steps.