[Question] Inconsistent and potentially swapped usage of pos_action_bounds and pos_action_threshold between FactoryEnv and ForgeEnv

[MpxThs]

Question

While studying the source code for the direct/factory and direct/forge environments, I noticed a significant discrepancy in how pos_action_bounds and pos_action_threshold (and their rotational counterparts) are applied to the policy actions. It seems these two variables might have been accidentally swapped in forge_env.py.

1. The Discrepancy in Action Scaling vs. Clipping

in FactoryEnv _apply_action():

• pos_thresholdis used as the action scale:pos_actions = self.actions[:, 0:3] * self.pos_threshold

• pos_action_bounds is used to clip the maximum workspace range:torch.clip(delta_pos, -self.cfg.ctrl.pos_action_bounds[0], ...)

in ForgeEnv _apply_action():

• pos_action_boundsis used as the action scale:`pos_actions = pos_actions @ torch.diag(torch.tensor(self.cfg.ctrl.pos_action_bounds, ...))
• pos_threshold is used to clip the delta step:torch.clip(self.delta_pos, -self.pos_threshold, self.pos_threshold)

2. The Contradiction with Config Values:in [FactoryEnvCfg](https://github.com/isaac-sim/IsaacLab/blob/main/source/isaaclab_tasks/isaaclab_tasks/direct/factory/factory_env_cfg.py) `CtrlCfg`,the default definitions are:

• pos_action_bounds = [0.05, 0.05, 0.05]
• pos_action_threshold = [0.02, 0.02, 0.02]
  Since bounds is much larger than threshold, the naming implies bounds should be the absolute safety limit, while threshold is the maximum single-step increment. The way they are used in ForgeEnv contradicts this intuition.

3. The Domain Randomization Paradox in ForgeEnv:
   in ForgeEnv._reset_idx(), domain randomization (noise) is added to self.pos_threshold

self.pos_threshold = forge_utils.get_random_prop_gains(
    self.pos_threshold, self.cfg.ctrl.pos_threshold_noise_level, ...
)

If pos_threshold is simply a hard clipping limit (as currently used in Forge_Env's clipping), applying noise to it seems physically meaningless. However, if pos_threshold was originally intended to be the action scale (as it is in FactoryEnv), then adding noise makes perfect sense.

Questions:

1. Was the swapping of these two variables in ForgeEnv intentional, or is it a bug?
2. If intentional, could you clarify the physical meaning of randomizing the clipping limit *pos_threshold in Forge)?

Thank you for your time and for maintaining this great repository!

Build Info

• Isaac Lab Version: 2.2.1
• Isaac Sim Version: [4.5]

[RandomOakForest]

Thank you for posting this. This is an interesting catch. Thank you for the detailed clarification. There hasn't been any changes to this in Isaac Lab 2.3.2. The team will review this and we will address it shortly.

[RandomOakForest]

Following up, the team is actively working to resolve this. Note that:

• This is a bug within the ForgeEnv; the variables were swapped erroneusly.
• The randomization should be on the scaling of the policy action, not the environment bounds