Automatic Waypoint Extraction (AWE)

[Project website] [Paper]

This repo contains the implementation of Automatic Waypoint Extraction (AWE): a plug-and-play module for selecting waypoints from demonstrations for performant behavioral cloning. This repo also includes instantiations of combining AWE with two state-of-the-art imitation learning methods, Diffusion Policy and Action Chunking with Transformers (ACT), and the respective benchmarking environments, RoboMimic and Bimanual Simulation Suite.

Given a set of demonstrations and an error threshold, extracting waypoints is as simple as:

pip install waypoint-extraction
import waypoint_extraction as awe
waypoints = awe.extract_waypoints(states, err_threshold)

If you encountered any issue, feel free to contact lucyshi (at) stanford (dot) edu

Installation

1. Clone this repository

   git clone git@github.com:lucys0/awe.git
   cd awe

2. Create a virtual environment

   conda create -n awe_venv python=3.9
   conda activate awe_venv

3. Install MuJoCo 2.1

   • Download the MuJoCo version 2.1 binaries for Linux or OSX.
   • Extract the downloaded mujoco210 directory into ~/.mujoco/mujoco210.

4. Install packages

   pip install -e .

RoboMimic

Set up the environment

# install robomimic
pip install -e robomimic/

# install robosuite
pip install -e robosuite/

Download data

# download unprocessed data from the robomimic benchmark
python robomimic/robomimic/scripts/download_datasets.py --tasks lift can square  

# download processed image data from diffusion policy (faster)
mkdir data && cd data
wget https://diffusion-policy.cs.columbia.edu/data/training/robomimic_image.zip
unzip robomimic_image.zip && rm -f robomimic_image.zip && cd ..

Usage

Please replace [TASK] with your desired task to train. [TASK]={lift, can, square}

• Convert delta actions to absolute actions

  python utils/robomimic_convert_action.py --dataset=robomimic/datasets/[TASK]/ph/low_dim.hdf5

• Save waypoints

  python utils/robomimic_save_waypoints.py --dataset=robomimic/datasets/[TASK]/ph/low_dim.hdf5 --err_threshold=0.005

• Replay waypoints (save 3 videos and 3D visualizations by default)

  mkdir video
  python example/robomimic_waypoint_replay.py --dataset=robomimic/datasets/[TASK]/ph/low_dim.hdf5 \
  --record_video --video_path video/[TASK]_waypoint.mp4 --task=[TASK] \
  --plot_3d --auto_waypoint --err_threshold=0.005

AWE + Diffusion Policy

Install Diffusion Policy

conda env update -f diffusion_policy/conda_environment.yaml

If the installation is too slow, consider using Mambaforge instead of the standard anaconda distribution, as recommended by the Diffusion Policy authors. That is:

mamba env create -f diffusion_policy/conda_environment.yaml

Train policy

python diffusion_policy/train.py --config-dir=config --config-name=waypoint_image_[TASK]_ph_diffusion_policy_transformer.yaml hydra.run.dir='data/outputs/${now:%Y.%m.%d}/${now:%H.%M.%S}_${name}_${task_name}'

Bimanual Simulation Suite

Set up the environment

conda env update -f act/conda_env.yaml

Download data

Please download scripted/human demo for simulated environments from here and save them in data/act/.

If you need real robot data, please contact Lucy Shi: lucyshi (at) stanford (dot) edu

Usage

Please replace [TASK] with your desired task to train. [TASK]={sim_transfer_cube_scripted, sim_insertion_scripted, sim_transfer_cube_human, sim_insertion_human}

• Visualize waypoints

  python example/act_waypoint.py --dataset=data/act/[TASK] --err_threshold=0.01 --plot_3d --end_idx=0 

• Save waypoints

  python example/act_waypoint.py --dataset=data/act/[TASK] --err_threshold=0.01 --save_waypoints 

AWE + ACT

Train policy

python act/imitate_episodes.py \
    --task_name [TASK] \
    --ckpt_dir data/outputs/act_ckpt/[TASK]_waypoint \
    --policy_class ACT --kl_weight 10 --chunk_size 50 --hidden_dim 512 --batch_size 8 --dim_feedforward 3200 \
    --num_epochs 8000  --lr 1e-5 \
    --seed 0 --temporal_agg --use_waypoint

For human datasets, set --kl_weight=80, as suggested by the ACT authors. To evaluate the policy, run the same command with --eval.

Citation

If you find our code useful for your research, please cite:

@inproceedings{shi2023waypointbased,
  title     = {Waypoint-Based Imitation Learning for Robotic Manipulation},
  author    = {Lucy Xiaoyang Shi and Archit Sharma and Tony Z. Zhao and Chelsea Finn},
  year      = {2023},
  booktitle = {Conference on Robot Learning}
}