Do MBPO agents not support recurrent neural networks for the environment model, the base off-policy agent, or both?

Since TD3, SAC, etc. agents support using recurrent layers by themselves, would using these recurrent base agents still not work with MBPO?

Could this limit be circumvented by using a custom training loop for the environment model and for the base agents?

[Naren Raman]

Thank you for your question. No, MBPO agents do not support recurrent networks for now as mentioned in the documentation. The custom training loop provides more flexibility. Yes, you should be able to use the custom training loop to create a custom MBPO agent with recurrent neural networks.

[Kundan Panta]

Thank you for your timely response. To confirm that recurrent networks are not supported for the base agents, in addition to the environment model, I tried combining the "Create MBPO Agent" and "Create TD3 Agent with Recurrent Neural Networks" examples and it indeed threw an error.