About integration with Gym

[pucciland95]

Hi,

We are trying to use this repo together with our ROS2 stack.
Since Mujoco is especially used to simulate for RL applications, I was wondering whether it was possible to use this in combination with Gym.
Is it possible to do so? How can I do it?

Cheers

[sangteak601]

Hi @pucciland95,

I haven't tested it with Gym yet, but it's something I really want to explore as well.

There are some potential issues I can think of at the moment:

1. Everything is running on the CPU.
2. Simulation states (e.g., the position of objects) are not exposed to the ROS2 interface, apart from joint states.
3. There is no way to reset the environment.

If you can test it out anyway and report issues or create a PR, that would be very helpful.

[pucciland95]

Hi, @sangteak601 and thank you for responding,

What do you think would be the best approach? I think we should need to forward the current state information from C/C++(Mujoco) to Python (Gym) somehow.
The great ChatGPT recommended to:

1. make some Python bindings
2. stream everything in ROS2 topic (for state) or services (for resetting)

I'm not sure, though, whether these changes should be in this repo or not since they have nothing to do with 'control' but rather with the state of the environment.

What are your opinions?

Cheers,

Niccolo