[Bug Report] Changed behavior of `set_external_force_and_torque()` from v2.3.1 to v2.3.2

[LoreMoretti]

Describe the bug

Release v2.3.2 deprecated set_external_force_and_torque() in favor of the composable wrench system.
But this causes some changed behavior in how external forces are applied in certain configuration, as I will try to document below.

Steps to reproduce

I have written an event with reset mode, where I am doing the following:

    asset.set_external_force_and_torque(
        forces, torques, env_ids=env_ids, body_ids=asset_cfg.body_ids, is_global=True
    )

with torques being 0.

The behavior is different if using IsaacLab v2.3.1 or v2.3.2. In particular, it seems that, in v2.3.2 release, the external forces are not being applied always in global frame.

-->

System Info

Describe the characteristic of your environment:

• Commit: 37ddf62
• Isaac Sim Version: [e.g. 5.0, this can be obtained by cat ${ISAACSIM_PATH}/VERSION]
• OS: Ubuntu 24.04
• GPU: RTX 5090
• CUDA: 13.0
• GPU Driver: 580.126.09

Additional context

I think the problem is related to this specific configuration of the external forces being applied at the reset, in global frame, and in permanent mode. Probably, the wrench composer was not taking into account this configuration, and the new API is not matching the behavior of v2.3.1 release.

It only affects the permanent wrench composer.

Checklist

• I have checked that there is no similar issue in the repo (required)
• I have checked that the issue is not in running Isaac Sim itself and is related to the repo

Acceptance Criteria

Add the criteria for which this task is considered done. If not known at issue creation time, you can add this once the issue is assigned.

• A way to replicate the same behavior of the deprecated API.

[LoreMoretti]

I think the problem is explained by the following example.

Let us consider a pure force (i.e., no torque) being applied at the reset (through an event) at a body frame (which can move during the simulation). For this force, we set is_global=True in the set_external_force_and_torque() method.

This case was working fine with the old APIs. For example, a force defined as [0,0,-10] was always pointing downwards in global frame, and was persistent for all the episode (from a reset instant to he following one).

Instead this is not true for the new APIs, where it seems that, if the body_frame is moving along the simulation, also the force changes is orientation (even though we are setting is_global=True).

I think this arises because:

1- At the reset, the permanent force is set at the reset in the wrench_composer by calling

IsaacLab/source/isaaclab/isaaclab/assets/rigid_object/rigid_object.py

Lines 461 to 468 in 37ddf62

	self._permanent_wrench_composer.set_forces_and_torques(
	forces=wp.from_torch(forces, dtype=wp.vec3f) if forces is not None else None,
	torques=wp.from_torch(torques, dtype=wp.vec3f) if torques is not None else None,
	positions=wp.from_torch(positions, dtype=wp.vec3f) if positions is not None else None,
	body_ids=body_ids,
	env_ids=env_ids,
	is_global=is_global,
	)

which gets the body position and orientation at this lines:

IsaacLab/source/isaaclab/isaaclab/utils/wrench_composer.py

Lines 292 to 297 in 37ddf62

	if not self._link_poses_updated:
	self._link_positions = wp.from_torch(self._get_link_position_fn().clone(), dtype=wp.vec3f)
	self._link_quaternions = wp.from_torch(
	convert_quat(self._get_link_quaternion_fn().clone(), to="xyzw"), dtype=wp.quatf
	)
	self._link_poses_updated = True

used later in:

IsaacLab/source/isaaclab/isaaclab/utils/wrench_composer.py

Lines 302 to 318 in 37ddf62

	wp.launch(
	set_forces_and_torques_at_position,
	dim=(env_ids.shape[0], body_ids.shape[0]),
	inputs=[
	env_ids,
	body_ids,
	forces,
	torques,
	positions,
	self._link_positions,
	self._link_quaternions,
	self._composed_force_b,
	self._composed_torque_b,
	is_global,
	],
	device=self.device,
	)

to project the force expressed in world frame to the local body frame in:

IsaacLab/source/isaaclab/isaaclab/utils/warp/kernels.py

Lines 458 to 459 in 37ddf62

	composed_forces_b[env_ids[tid_env], body_ids[tid_body]] = cast_force_to_link_frame(
	forces[tid_env, tid_body], link_quaternions[env_ids[tid_env], body_ids[tid_body]], is_global

which are then stored in composed_forces_b warp array and, omitting some function call, eventually saved also in the wrench composer instance attribute self._composed_force_b_torch.

2- From now on, for all the following simulation steps until the next reset event, this force is "written to the simulation" by:

IsaacLab/source/isaaclab/isaaclab/assets/rigid_object/rigid_object.py

Lines 160 to 167 in 37ddf62

	# Apply permanent wrench to the simulation
	self.root_physx_view.apply_forces_and_torques_at_position(
	force_data=self._permanent_wrench_composer.composed_force_as_torch.view(-1, 3),
	torque_data=self._permanent_wrench_composer.composed_torque_as_torch.view(-1, 3),
	position_data=None,
	indices=self._ALL_INDICES,
	is_global=False,
	)

This function applies the projected force (from world to body local frame) in the body local frame (is_global=False).
This breaks if the body local frame is moving during the simulation, because no additional projection is computed and applied: the force is expressed in the old local frame and applied in the new local frame (without any needed projection).

If for example, the body local frame was first aligned with the world frame and then, during the simulation, rotates of 180[deg] around the x-axis, a pure z-force (in world frame) ends up being applied in the opposite direction.

This behavior was not happening in the release v2.3.1, where the same force would have had a constant direction (in world representation) until the next reset event.

[RandomOakForest]

Thank you for posting this and following up with the example. The team will review this.