Add policy/act_aloha_real.yaml + env/act_real.yaml

[Cadene]

What this does

• Add support for training on Aloha
• Add multithreaded multiprocessing during recording
  • Before the main process was creating image writers, but this was causing slow fps during record inference on Aloha because of the 4 cameras
  • Now, the main process creates a subprocess that creates all the image writers

How it was tested

• Recorded a dataset and verified no out of RAM during compute stats
• Started training

- Ran inference and verified fps is fast

How to checkout & try? (for the reviewer)

Note: Use DATA_DIR=data for local dataset

python lerobot/scripts/control_robot.py record \
  --robot-path lerobot/configs/robot/aloha.yaml \
  --fps 30 \
  --root data \
  --repo-id cadene/aloha_transfer_cube \
  --num-episodes 50 \
  --run-compute-stats 1 \
  --warmup-time-s 5 \
  --episode-time-s 360 \
  --reset-time-s 360 \
  --force-override 0

Started training

python lerobot/scripts/train.py \
  dataset_repo_id=cadene/aloha_lace_shoe \
  policy=act_aloha_real \
  env=aloha_real \
  hydra.run.dir=outputs/train/act_aloha_lace_shoe \
  hydra.job.name=act_aloha_lace_shoe \
  device=cuda \
  wandb.enable=true

Ran inference

python lerobot/scripts/control_robot.py record \
  --robot-path lerobot/configs/robot/aloha.yaml \
  --fps 30 \
  --root data \
  --repo-id cadene/eval_act_transfer_cube_080000 \
  --num-episodes 40 \
  --run-compute-stats 0 \
  --warmup-time-s 5 \
  --episode-time-s 360 \
  --reset-time-s 360 \
  --force-override 0 \
  -p outputs/train/act_aloha_transfer_cube/checkpoints/080000/pretrained_model

[aliberts]

LGTM (shallow review)
Let's improve on the control_robot.py in a PR soon so that it's more readable ;)