Add RLBench

[eugeneteoh]

What this does

Add RLBench as a benchmark.
https://github.com/stepjam/RLBench

How it was tested

How to checkout & try? (for the reviewer)

conda create -y -n lerobot python=3.10 && conda activate lerobot
conda env config vars set COPPELIASIM_ROOT=/path/to/coppeliasim
conda env config vars set LD_LIBRARTY_PATH=$LD_LIBRARY_PATH:$COPPELIASIM_ROOT QT_QPA_PLATFORM_PLUGIN_PATH=$COPPELIASIM_ROOT 
conda activate lerobot
pip install -e ".[rlbench]"
# First, generate raw rlbench data
rlbench-generate-data --save_path=data/raw/rlbench --tasks=reach_target --episodes_per_task=100 --variations=1 --processes=4
# Push dataset to huggingface hub
python lerobot/scripts/push_dataset_to_hub.py --raw-dir data/raw/rlbench/open_drawer/variation0 --raw-format rlbench --repo-id eugeneteoh/rlbench_reach_target_variation_0 --local-dir data/eugeneteoh/rlbench_open_drawer_variation_0 --push-to-hub=1
# Run training
python lerobot/scripts/train.py policy=act_rlbench env=rlbench

[alexander-soare]

I suppose you plan to revert all of the changes in this file?

[eugeneteoh]

Yes

[eugeneteoh]

I'm actually not sure how to deal with the config for rlbench. For now, I have added a separate config file.

[alexander-soare]

So why these changes?

[eugeneteoh]

Main reason is removing this line:
https://github.com/huggingface/lerobot/pull/290/files/e7d4901301a35e18dcfcf6e4f04b68f84d2a929a#diff-0f7c1a6977da530f85035451866c98b1938bee512d68bfd4febfb1a8e367114eL30

Previously, there was an assumption that all package names of the environments were prefixed by gym_. In rlbench, this is not the case. I could either keep this assumption and have a special case for rlbench, or remove the assumption. I thought the latter was better.

[alexander-soare]

Thanks @eugeneteoh, this is about what I had in mind for supporting multiple "state" vectors. I've reviewed the modeling code and will leave the dataset stuff to @Cadene

[alexander-soare]

Yeah so as I was mentioning before. Don't you already have this set up as an instance attribute at this point?

[zwbx]

Hi, when will this feature be merged? Are there any ways to run act on rlbench?

[github-actions]

This PR has been automatically marked as stale because it has not had recent activity (6 months). It will be closed if no further activity occurs. Thank you for your contributions.

[pyun-ram]

Hi what about the recent update in this PR? I wonder ask will this feature be merged into lerobot recently? :)