Custom RL Algorithm

Hi, I have create a variations on SAC that I would like to test on Isaac Gym. Those anyone know how to test your own algorithm in Isaac Gym?

Hi @sesem738

Could you, please, extend the description a bit to try to provide better help? For example, do you already have the learning algorithm implemented, or a whole implementation that you want to connect with Isaac Gym, or does the modification fall on the neural networks, the replay memory, etc…?

In any case, I encourage you to try the skrl RL library with full support for Isaac Gym previews, Isaac Orbit, Omniverse Isaac Gym, and other environment interfaces. It will be easy for you to create new or modify the existing algorithms, taking advantage of existing modules :)

I took a look at SRKL and it looks to be what I am looking for. I am trying to implement a recurrent SAC algorithm that I would later use on the real robot. I have been running into issue trying to use it. Please, do you have any good examples/repositories that I could reference?

Hi @sesem738

Can you elaborate on the issues you are facing?

For recurrent implementation, you can find some scripts in the skrl documentation examples folder.

The _rnn_initial_states become empty after the first random act

Hi @sesem738

Can you please share the main script or show a snippet of the defined models?