This repository is the implementation code of the paper “Co-GAIL: Learning Diverse Strategies for Human-Robot Collaboration”(arXiv, Project, Video) by Wang et al. at Stanford Vision and Learning Lab. In this repo, we provide our full implementation code of training and evaluation.


  • python 3.6+

conda create -n cogail python=3.6
conda activate cogail
  • iGibson 1.0 variant version for co-gail. For more details of iGibson installation please refer to Link

git clone --recursive
cd iGibson
git checkout cogail
python -m pip install -e .

Please also download the assets of iGibson (models of the objects, 3D scenes, etc.) follow the instruction. The data should be located at your_installation_path/igibson/data/. After downloaded the dataset, copy the modified robot and humanoid mesh file to this location as follows

cd urdfs
cp fetch.urdf your_installation_path/igibson/data/assets/models/fetch/.
cp camera.urdf your_installation_path/igibson/data/assets/models/grippers/basic_gripper/.
cp -r humanoid_hri your_installation_path/igibson/data/assets/models/.
  • other requirements

cd cogail
python -m pip install -r requirements.txt


You can download the collected human-human collaboration demonstrations for Link. The demos for cogail_exp1_2dfq is collected by a pair of joysticks on an xbox controller. The demos for cogail_exp2_handover and cogail_exp3_seqmanip are collected with two phones on the teleoperation system RoboTurk. After downloaded the file, simply unzip them at cogail/ as follows

mv dataset your_installation_path/cogail/dataset


There are three environments (cogail_exp1_2dfq, cogail_exp2_handover, cogail_exp3_seqmanip) implemented in this work. Please specify the choice of environment with --env-name

python scripts/ --env-name [cogail_exp1_2dfq / cogail_exp2_handover / cogail_exp3_seqmanip]


Evaluation on unseen human demos (replay evaluation):

python scripts/ --env-name [cogail_exp1_2dfq / cogail_exp2_handover / cogail_exp3_seqmanip]

Trained Checkpoints

You can download the trained checkpoints for all three environments from Link.


The cogail_exp1_2dfq is implemented with Pygame. The cogail_exp2_handover and cogail_exp3_seqmanip are implemented in iGibson v1.0.

The demos for robot manipulation in iGibson is collected with RoboTurk.

Code is based on the PyTorch GAIL implementation by ikostrikov (


Please cite Co-GAIL if you use this repository in your publications:

  title={Co-GAIL: Learning Diverse Strategies for Human-Robot Collaboration},
  author={Wang, Chen and P{\'e}rez-D'Arpino, Claudia and Xu, Danfei and Fei-Fei, Li and Liu, C Karen and Savarese, Silvio},
  journal={arXiv preprint arXiv:2108.06038},


Licensed under the MIT License


View Github