Skip to content

Latest commit

 

History

History
98 lines (86 loc) · 4.77 KB

Instructions.md

File metadata and controls

98 lines (86 loc) · 4.77 KB

Installation Instructions

  • Clone the repository and create a conda environment using the provided conda_env.yml file.
conda env create -f conda_env.yml
  • Activate the environment using conda activate baku.
  • To install LIBERO, follow the instructions in the LIBERO repository.
  • To install Meta-World, install the version added to this repository as a submodule.
git submodule update --init --recursive
cd Metaworld
pip install -e .

Instructions for Real Robot Experiments

  • For running experiments on an xArm robot, install the xArm-Python-SDK using the instructions provided in the repository.
  • Install the xarm environment using the following command.
cd xarm_env
pip install -e .
  • For teleoperating the robot using Open Teach, use the version of Open Teach added to this repository as a submodule. Move inside the directory using cd Open-Teach and install the package using the instructions provided in the official Open-Teach repository.
    • Instructions for robot teleoperation and data collection are provided here.
  • To run the experiments, use the following command.

Instructions for Demonstrations

  • For LIBERO, download the demonstrations using instructions provided in the LIBERO repository.
    • Convert these demonstrations into a .pkl format using the following commands.
    cd baku/data_generation
python generate_libero.py
    • Make sure to change the DATASET_PATH in baku/data_generation/generate_libero.py to the path where the demonstrations are stored.
    • The script stores the demonstration pkl files in path/to/repo/expert_demos/libero/. Set root_dir in cfg/config.yaml to path/to/repo.
  • To access the datasets for Meta-World, DMControl, and the real world xArm Kitchen, please send an email to [email protected].

Train BAKU

  • Remember to set root_dir in cfg/config.yaml to path/to/repo.
  • To train BAKU on LIBERO-90, use the following command.
python train.py agent=baku suite=libero dataloader=libero suite/task=libero_90 suite.hidden_dim=256
  • To train BAKU on Meta-World, use the following command.
python train.py agent=baku suite=metaworld dataloader=metaworld suite.hidden_dim=256 use_proprio=false
  • To train BAKU on DMControl, use the following command.
python train.py agent=baku suite=dmc dataloader=dmc suite.hidden_dim=256 obs_type=features use_proprio=false

Evaluate BAKU

  • We provide the weights for BAKU trained on the simulated benchmarks here. Please download this weights to evaluate BAKU on the benchmarks.
  • We provide 2 variants of evaluation scripts - one for sequential evaluation (eval.py) and one for parallel evaluation (eval_ray.py to use which you will have to install Ray).
  • Before evaluating, make sure to set the root_dir in cfg/config_eval.yaml to path/to/repo.
  • For evaluation on LIBERO-90, use one of the the following commands.
python eval.py agent=baku suite=libero dataloader=libero suite/task=libero_90 suite.hidden_dim=256 bc_weight=/path/to/weight

python eval_ray.py agent=baku suite=libero dataloader=libero suite/task=libero_90 suite.hidden_dim=256 bc_weight=/path/to/weight

Follow the same pattern for evaluation on Meta-World and DMControl.

Train Baselines

MT-ACT

  • To train MT-ACT on LIBERO-90, use the following command.
python train.py agent=mtact suite=libero dataloader=libero suite/task=libero_90
  • To train MT-ACT on Meta-World, use the following command.
python train.py agent=mtact suite=metaworld dataloader=metaworld use_proprio=false
  • To train MT-ACT on DMControl, use the following command.
python train.py agent=baku suite=dmc dataloader=dmc obs_type=features use_proprio=false

RT-1

  • To train RT-1 on LIBERO-90, use the following command.
python train.py agent=rt1 suite=libero dataloader=libero suite/task=libero_90 suite.hidden_dim=512 suite.history=true suite.history_len=6 temporal_agg=false
  • To train RT-1 on Meta-World, use the following command.
python train.py agent=rt1 suite=metaworld dataloader=metaworld suite.hidden_dim=512 use_proprio=false suite.history=true suite.history_len=6 temporal_agg=false
  • To train RT-1 on DMControl, use the following command.
python train.py agent=rt1 suite=dmc dataloader=dmc suite.hidden_dim=512 obs_type=features use_proprio=false suite.history=true suite.history_len=6 temporal_agg=false