[Question] Cannot import trained policy into navigation RL

[H-Hisamichi]

Hello everyone,

I am working on training a robot for a navigation task using a custom policy.
I used the default navigation code as a base. However, I encountered the following error when trying to use the trained policy.

Error:
RuntimeError: PytorchStreamReader failed locating file constants.pkl: file not found

When performing RL using the default code, agent.pkl and env.pkl are saved in the params folder, and the .pt file is stored in the same directory as the params folder. However, there is no mention of a constants.pkl file, so it is unclear what the error is referring to.

Does this file need to be generated manually, or is there a required preprocessing step to convert the trained policy before using it in the navigation RL framework?

Any help would be appreciated. Thank you!

[H-Hisamichi]

Additional note:
Upon further investigation, I found that on_policy_runner.py saves the model in state_dict format, while PreTrainedPolicyAction tries to load the model in TorchScript format.
I suspect this mismatch is the cause of the error.
Does Isaac Lab provide any settings or tools for specifying or converting the model format?

[H-Hisamichi]

Hello everyone,

This problem was solved.
Running play.py generated a model in TorchScript format.