I am working with Isaac Sim / Lab and encountered some difficulties when deploying pretrained reinforcement learning policies.
Context:
-
My policies are trained in Isaac Lab using default PPO algorithm and SKRL framework (training script path is:
IsaacLab\source\isaaclab_tasks\isaaclab_tasks\manager_based\locomotion\velocity\velocity_env_cfg.py -
For deployment, I use a customized version of the example script:
isaacsim\exts\isaacsim.robot.policy.examples\isaacsim\robot\policy\examples\robots\h1\H1FlatTerrainPolicy
Problems:
-
Weights format mismatch:
-
Isaac Lab saves weights .pt file as Python dictionary (dict with policy, value, and optimizer keys)
-
Isaac Sim’s policy loader expects a
torch.jit._script.RecursiveScriptModuleobject to be loaded withtorch.jit.load -
Attempting
torch.jit.loadleads to errors like:RuntimeError: PytorchStreamReader failed locating file constants.pkl: file not found
-
-
Observation vector ambiguity**:**
-
The
_compute_observationfunction must manually define the observation vector -
However, in Isaac Lab manager-based training setup, the observation vector is abstracted away, making it unclear what the exact observation length / structure is.
-
Questions:
-
Is there an official or recommended workflow to convert SKRL-trained policies into TorchScript modules (
torch.jit._script) compatible with Isaac Sim? -
How can we extract the exact observation vector structure (dimension and components) from an Isaac Lab manager-based setup to ensure
_compute_observationmatches deployment requirements?
All in all, should we expect Isaac Lab’s pretrained policies (from SKRL or RSL-RL) to be deployable out-of-the-box in Isaac Sim, or is manual conversion always required?
Since, I am adapting the official example scripts rather than building from scratch, I assumed deployment would be seamless. Therefore, any guidance on the intended workflow would be greatly appreciated.
Isaac Sim Version
5.0.0
Operating System
Windows 10
GPU Information
- Model: GeForce RTX 5070
- Driver Version: 577.00
- CUDA Version: 12.9