Fix Isaac Lab force clear shapes #74
Conversation
|
which version of IsaacLab are you using? MimicKit has been tested with |
| positions=None, env_ids=None, | ||
| body_ids=None, is_global=True) | ||
| num_envs = self.get_num_envs() | ||
| for obj_id, obj in enumerate(self._objs): |
There was a problem hiding this comment.
I think a much easier fix would be just to change this line to
if(getattr(obj, 'has_external_wrentch', False)):
then you don't need any other changes.
Articulation still has has_external_wrentch but unfortunately it was only set after you call set_external_force_and_torque.
There was a problem hiding this comment.
I tested this change with latest Isaaclab from uv pip install isaaclab and it works.
There was a problem hiding this comment.
@ustcscgyer thanks for the suggested change! That looks good to me, simple and elegant. Lets go with this fix instead. Would you like to submit a separate PR with your fix?
@Michael-Jetson does this fix the issues you were encountering on your side?
3 participants
Find some bug when I try to train in my pc,and I fix it by codex-5.2,env info is:
cuda 12.8
isaacsim 5.1.0
RTX 5080
Ubuntu 24.03
python 3.11
torch 2.7.0
Root cause: some Isaac Lab builds no longer expose Articulation.has_external_wrench, so _clear_forces hit AttributeError. After a quick workaround, the force-
clearing path passed a wrongly shaped tensor (e.g., [1, 3]), causing set_forces_and_torques_at_position to fail because it expects 2D/3D force arrays.
Changes in mimickit/engines/isaac_lab_engine.py:
with correct shapes based on object type: articulated uses [num_envs, num_bodies, 3], rigid uses [num_envs, 3], then resets the flag.
This prevents API-compat crashes and ensures force clearing uses valid tensor dimensions.