Motivation

The wrapper to use DMCS with gym has an interesting feature that allows you to learn directly with the rendering (pixels), and not with the state. I don't see this feature integrated in envpool. Is it planned?

Solution

Could be an additional env kwargs:

envpool.make_dm("BallInCupCatch-v1", num_envs=1, from_pixels=True)

Or register those envs:

envpool.make_dm("BallInCupCatchImage-v1", num_envs=1)  # or BallInCupCatchFromPixels-v1, doesn't matter

Alternatives

I don't any workaround to use DeepMind control suite from pixel with envpool in the current state.

Additional context

From the paper

We also benchmarked a variant of D4PG that learns directly from pixel-only input, using 84 × 84 RGB frames from the 0th camera.

Checklist

• I have checked that there is no similar issue in the repo (required)