Generalizable Domain Adaptation for Sim-and-Real Policy Co-Training

This repository provides the implementation for Generalizable Domain Adaptation for Sim-and-Real Policy Co-Training.
Our implementation builds on MimicGen, Robosuite, and Robomimic. To keep the codebase compact, we include a modified version of Robomimic in this repo. You must also install the official versions of Robosuite and MimicGen (following the installation instructions below).

Installation

Install MimicGen

conda create -n ot-sim2real python=3.8
conda activate ot-sim2real
cd <PATH_TO_YOUR_INSTALL_DIRECTORY>
git clone https://github.com/NVlabs/mimicgen.git
cd mimicgen
pip install -e .

Install Robosuite

cd <PATH_TO_YOUR_INSTALL_DIRECTORY>
git clone https://github.com/ARISE-Initiative/robosuite.git
cd robosuite
git checkout b9d8d3de5e3dfd1724f4a0e6555246c460407daa
pip install -e .

Install modified version of robomimic

cd <PATH_TO_YOUR_INSTALL_DIRECTORY>
git clone
cd ot-sim2real
pip install -e .

Install robosuite-model-zoo

cd <PATH_TO_YOUR_INSTALL_DIRECTORY>
git clone https://github.com/touristCheng/model-zoo-mugs-only.git
cd model-zoo-mugs-only
pip install -e .

Download Datasets

Specify the desired dataset directory in robomimic/scripts/ot_sim2real/download_datasets.py, and run the script

cd <PATH_TO_YOUR_INSTALL_DIRECTORY>
cd ot-sim2real
python robomimic/scripts/ot_sim2real/download_datasets.py

Generate configs

Our proposed method is denoted as ot-sim2real.
We also include four baselines:

• MMD
• cotrain
• source_only
• target_only

The experiment configs can be generated using robomimic/scripts/ot_sim2real/generate_exp_configs.py. Please specify the DATA_ROOT, OUTPUT_DIR, and GENERATED_CONFIG_DIR, and then run the script

cd <PATH_TO_YOUR_INSTALL_DIRECTORY>
cd ot-sim2real
python robomimic/scripts/ot_sim2real/generate_exp_configs.py

Run training

For ot-sim2real and MMD, the training script is robomimic/scripts/ot_sim2real/ot_train.py.

# for "xxx_ot-sim2real.json" and "xxx_MMD.json"
cd <PATH_TO_YOUR_INSTALL_DIRECTORY>
cd ot-sim2real
python robomimic/scripts/ot_sim2real/ot_train.py --config path/to/config.json

For cotrain, source_only and target_only, the training script is robomimic/scripts/train.py.

# for "xxx_cotrain.json", "xxx_source_only.json" and "xxx_target_only.json"
cd <PATH_TO_YOUR_INSTALL_DIRECTORY>
cd ot-sim2real
python robomimic/scripts/train.py --config path/to/config.json