skrl agent and torch nn implementations are yielding different results

[hsjung02]

Hello, thanks for your great work.

I am working on skrl with Isaac Lab, and trying to convert the skrl model to pytorch nn.module-based implementation.
I need to implement the model using pytorch nn.module because I am willing to run the neural network on cpp(and pytorch supports cpp implementation via libtorch).

To this end, I parsed the network from agent.yaml and loaded parameters(weights and biases) from best_agent.pt

However, as I implement the nn.module and tested using zero input(obs with all zeros), the skrl agent and my custom model showed different outputs.

I reviewed my code and I cannot find problem on myself. Can anybody help??

I attached my full code for reproduction.

skrl_torch_comparison.zip

I was using

• skrl = 1.4.3
• torch = 2.7.0+cu128

[hsjung02]

This was because I did not use the scaler.
After implementing scaler by parsing running_mean and running_variance, both work the same!