Issues deploying pretrained policies in Isaac Sim

[PsorTheDoctor]

When using SKRL inside Isaac Lab to train RL policies, the program saves weights .pt file as a Python dict (with policy, value, and optimizer keys). It can be further loaded with torch.load function.

However, Isaac Sim's policy loader expects a torch.jit._script.RecursiveScriptModule object to be loaded with torch.jit.load and attempting torch.jit.load leads to errors like:

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

Is there any way to convert between these file content formats? This may be the crucial issue regarding usage of SKRL inside Isaac Lab / Sim, so I posted the original thread also on their repo if you find this useful: isaac-sim/IsaacLab#3697

[Toni-SM]

Hi @PsorTheDoctor

Currently, there is no implementation to export the trained policies as PyTorch's ScriptModule using torch.jit.save or ONNX.
This implementation has been on the roadmap for a while now. Unfortunately, I hadn't had time to carry it out.

While I understand such feature is crucial for sim-to-real, it probably won't be in the upcoming 2.0.0 release.