Local Chunk Normalization

Install

micromamba create -n dynamic_chunk python=3.12
micromamba activate dynamic_chunk
pip install git+https://github.com/huggingface/lerobot.git@0cf864870cf29f4738d3ade893e6fd13fbd7cdb5
pip install -e .

# download pushT data
python scripts/download.py

Train

# original DP-Unet model
python scripts/train_dp_unet_pushT.py  # need to change into your own wandb project if needed


# Add weighted loss, weights are calculated according to the local chunk norm
python scripts/train_dp_unet_pushT_lcn_weighted.py 

Eval

# eval the ckpt directory, then will calculate all saved ckpts in the dir
python scripts/eval_dp_unet_pushT.py --ckpt_dir runs/train_dp_unet_pushT/bs64_obs2_exe8_pre16_stdnorm_delta

# eval individual ckpt 
python scripts/eval_dp_unet_pushT.py  --ckpt_path runs/train_dp_unet_pushT/bs64_obs2_exe8_pre16_stdnorm_delta/ckpt_785000.pth

Task

• Finish original DP-Unet on PushT
• Finish weighted loss DP-Unet on PushT.
• [] Finish experiments on LIBERO