Olaf

Interactive Robot Learning from Verbal Correction

Running it locally

Setup:

conda create -n lflf python=3.8
cd LfLF
pip install -e .
pip install -r requirements.txt

downloading dataset:

python lflf/helper_scripts/download_file.py 

parallel relabeling:

change --mode to be one of {edit_action, return_action, llm_critic}

python lflf/scripts/relabeling_actions_parallel.py --original_dataset data/round01_square.hdf5 --relabeled_dataset data/round01_square_edit_action_long.hdf5 --mode edit_action --candidate_strategy onedim --num_actions 8 --gpt_model gpt-35-turbo --language_correction data/language_correction_long.npy --temperature 0.5 --num_threads 100

non-parallel relabeling:

Specify the trajectory with --ep.

change --mode to be one of {edit_action, return_action, llm_critic}

python lflf/scripts/relabeling_actions_single_traj.py --original_dataset data/round01_square.hdf5 --relabeled_dataset data/round01_square_edit_action_long.hdf5 --mode edit_action --candidate_strategy onedim --num_actions 8 --gpt_model gpt-35-turbo --language_correction data/language_correction_long.npy --temperature 0.5 --ep demo_84