[Question] What’s the recommended way to access randomized external forces and torques applied to assets for manager-based workflow?

[diracdelta7]

Question
First, thank you for the fantastic work!

I'm currently working on legged locomotion using a teacher-student architecture. In my setup, I define external_forces_and_torques in velocity_env_cfg.py as follows:

# In velocity_env_cfg.py
external_forces_and_torques = ObsTerm(
    func=mdp.external_forces_and_torques,
    params={"sensor_cfg": SceneEntityCfg("robot", body_names=".*base")},
)

And in mdp.observation.py, I define the function:

# In mdp.observation.py
def external_forces_and_torques(env: ManagerBasedRLEnv, asset_cfg: SceneEntityCfg) -> torch.Tensor:
    """External forces and torques applied to the body."""
    asset: Articulation = env.scene[asset_cfg.name]

    external_force_b = asset._external_force_b[:, asset_cfg.body_ids, :].squeeze(1)
    external_torque_b = asset._external_torque_b[:, asset_cfg.body_ids, :].squeeze(1)
    return torch.cat((external_force_b, external_torque_b), dim=1)

The issue is that asset._external_force_b and asset._external_torque_b are protected members of the Articulation class, and accessing them triggers a warning. Is there a more appropriate way to access these variables?

Additionally, in a teacher-student setup, it’s common to design events that randomize the environment or apply specific changes, such as forces, torques, or friction adjustments. For events like these, we may want to capture information about these randomizations as privileged observations for the teacher. However, accessing this information directly isn’t always feasible (e.g., asset._external_force_b), and in some cases, it can be time-consuming to compute (e.g., retrieving static friction via asset.root_physx_view.get_material_properties() requires CPU resources). Is there a recommended approach in IsaacLab for handling this efficiently?