[Bug]: Manipulation Examples can't run on v0.3.4

[EZ-Blue]

Bug Description

Hello,

I am trying to run the manipulation example script grasp_train.py and grasp_eval.py. However, when running the scripts on the lastest version of Genesis (v0.3.4), I am getting this exception error:

[Genesis] [14:31:31] [ERROR] GenesisException: Impossible to set env-specific pos for fixed links with at least one geometry.

When I downgrade to v0.3.3, I am able to run the grasp_train.py and grasp_eval.py with no exceptions, and I can train and evaluate the trained model with no issues.

I believe the exception error comes from commit ad6ef69 "[BUG FIX] Raise exception in case of invalid set entity pos/quat. (#1847)".

In the grasp_train.py script, on line 134 in the env_cfg "box_fixed" is set to True. This box config is used in grasp_env.py on line 76 for the manipulation object in the example. The object position is randomized for the training, on line 167 where self.object.set_pos(random_pos, envs_idx=envs_idx) is called. This is where the exception error occurs.

The changes in commit ad6ef69 raise an exception error when trying to set the pos of a fixed link, which causes the error in the manipulation examples.

Changing the box_fixed config to False also does not solve this error, as after the initialization successfully finishes, there will be a Runtime error upon the first environment reset: [Genesis] [14:53:20] [ERROR] RuntimeError: The size of tensor a (4) must match the size of tensor b (2048) at non-singleton dimension 0.

I believe this error is related to the same commit ad6ef69, because the error arises at line 1624 set_all_envs = (torch.sort(envs_idx).values == self._scene._envs_idx).all().

Let me know if there's anything else I can provide to help with reproducing or examining this issue!

Steps to Reproduce

Run grasp_train.py on Genesis v0.3.4, commit version ad6ef69.

Expected Behavior

The manipulation training script should be able to run without raising the exception error for setting the position of the box object to the randomly generated position.

Screenshots/Videos

No response

Relevant log output

Environment

• OS: Ubuntu 22.04
• GPU: RTX 5080
• NVIDIA Driver Version: 570.195.03
• CUDA 12.8

Release version or Commit ID

v0.3.4

Commit ad6ef69

Additional Context

No response

[Kashu7100]

Does this PR fix the issue? #1869
If you could test for us, that will be very helpful.

[duburcqa]

A patch was already merged to fix this issues several days ago. Please update to the latest main branch and let us know if it works. If not, I will reopen. Thanks !

[hckang-dev]

@duburcqa The patch doesn't work😢

[EZ-Blue]

Hello - sorry for the delay in response I was on break. I can also confirm that the patch does not work for the grasp_train.py, on v0.3.4 or v0.3.5. However, setting the "box_fixed" argument in the env_cfg to "False" allows you to bypass this issue and I no longer get the environment mismatch error.

[phuicy]

Setting the "box_fixed" argument in the env_cfg to "False" also works for me to resolve this issue. But then I find the default params lead to the RL to fail; Losses look good, but the robot is just slapping the table, the behavior cloning does nicely clone the robot slapping the table :)

[duburcqa]

OK I will fix it.

[duburcqa]

Fixed by #1998

[phuicy]

Thank you for your fast response. I pulled main fresh, and it now seems to be working.