Is there a way to deploy policy models to devices that cannot install ray

Thanks to rllib for this great library, I trained a ppo model with rllib, and I want to deploy this model to jetson nano, but it is too difficult to run rllib on jetson nano, is there a way to run ppo model directly without installing ray? thanks for help!

Hi @QiZhen_Chen ,

Out of the box, this is difficult. Our own abstractions obviously require Ray, so you will never be able to instantiate a Policy just like that without installing Ray. Also not in the future. We are taking steps towards easier productization with the Connectors project and other ongoing efforts.

What you are trying to do, you can only achieve by checkpointing the policy while training with RLLib and building a light-weight wrapper around the neural network yourself without Ray. You’d have to implement preprocessing and an action distribution yourself to match what RLlib does.