feat(act): add a config option for adding a bottleneck layer after encoding joint states

[villekuosmanen]

What this does

Adds a config option to introduce a bottleneck layer to ACT between the joint states and the rest of the model. Suggested by @Cadene on Discord.

How it was tested

I trained a model using this option. Not 100% sure if it was better or worse but a good tool to have.

Also tested not having the flag in the config, it defaults to false.

How to checkout & try? (for the reviewer)

You need to have an ALOHA dataset to try it out of the box. Otherwise could change act_koch_real to turn on the flag, then train a model.

DATA_DIR=data python lerobot/scripts/train.py \
  dataset_repo_id=${HF_USER}/koch_test \
  policy=act_koch_real \
  env=koch_real \
  hydra.run.dir=outputs/train/act_koch_test \
  hydra.job.name=act_koch_test \
  device=cuda \
  wandb.enable=true

[alexander-soare]

@villekuosmanen thanks for looking into this. Sorry, but we've discussed this in the LeRobot team and decided we don't want to add features if we don't have a strong indication that they add value. I think to get this over the line we'd have to prove it works (which would be a nice result worth publishing at some level).