ROS2 Deployment rl_games Training

[AbbassIbrahim01]

I am currently facing an issue and need clarification: if I train a policy using rl_games (Forge or Factory) and the output is a .pth checkpoint file, can this model be deployed directly on a real robot through ROS2, or does the .pth file need to be converted into another format—such as TorchScript or ONNX—before it can be used for Sim-to-Real deployment?

[ashwinvkNV]

You can do both. The easiest way is to write a custom rl_games ros inference node(rsl_rl example).

But for better inference performance, you can also export it as an onnx and use tensor_rt.