ricky-zhu/cycle_dynamics

[ICLR2021, Oral] Learning Cross-Domain Correspondence for Control with Dynamics Cycle-Consistency

Cycle-Dynamics

This repository contains the official implementation for Cycle-Dynamics introduced in the following paper:

Learning Cross-Domain Correspondence for Control with Dynamics Cycle-Consistency

Qiang Zhang, Tete Xiao, Alexei A. Efros, Lerrel Pinto, Xiaolong Wang

The project page with video is at https://sjtuzq.github.io/cycle_dynamics.html.

Citation

If you find our work useful in your research, please cite:

@article{zhang2020learning,
  title={Learning Cross-Domain Correspondence for Control with Dynamics Cycle-Consistency},
  author={Zhang, Qiang and Xiao, Tete and Efros, Alexei A and Pinto, Lerrel and Wang, Xiaolong},
  journal={arXiv preprint arXiv:2012.09811},
  year={2020}
}

Environment

• Python 3.6
• PyTorch 1.0 or higher, with NVIDIA CUDA Support
• Other required python packages are specified by requirements.txt.

Prerequisites

• Install Mujoco and other required packages.
• Clone this repository.
• Create the log folder './logs/cross_physics', './logs/cross_modality' and './logs/cross_morphology'.

Reproducing Cross-physics Experiment

1. Train the policy network, and collect the source domain data. Or you can download the pretrained policy weight here. Then place it in './logs/cross_physics/', unzip this file and collect the source domian data.

cd cross_physics/base_train_test/td3_solver
python train.py --env 'HalfCheetah-v2'
cd ../../cycle_transfer
python collect_data.py --data_type 'base' --data_id 1

2.Modify the xml file to change the environment, then collect data

cd cross_physics/base_train_test/mujoco_xml
python modify_xml.py --env 'half_cheetah' --phy 'arma3'
cd ../../cycle_transfer
python collect_data.py --data_type 'arma3' --data_id 1

3.Train the model to learn the correspondence

cd cross_physics/cycle_transfer
python forwardexp.py --data_type1 'base' --data_id1 1 --data_type2 'arma3' --data_id2 1

Reproducing Cross-modality Experiment

1. Train the policy network and collect the source domain data. Or you can download the pretrained policy weight here. Then place it in './logs/cross_modality/', unzip this file and collect the source domian data.

cd cross_modality/mujoco_exp/base_train_test/td3_solver
python train.py --env 'HalfCheetah-v2'

2.Collect the unpaired data for both domains.

cd cross_modality/mujoco_exp/cycle_transfer/
python collect_data.py --env 'HalfCheetah-v2' --data_type 'base' --data_id 1

3.Train the model to learn the correspondence.

cd cross_modality/mujoco_exp/cycle_transfer/
python cycleexp.py --data_type1 'base' --data_id1 1 --data_type2 'base' --data_id2 1

Reproducing Cross-morphology Experiment

1. Train the policy network and collect the source domain data. Or you can download the pretrained policy weight here. Then place it in './logs/cross_morphology/', unzip this file and collect the source domian data.

cd cross_morphology/base_train_test/td3_solver
python train.py --env 'HalfCheetah-v2'
cd ../../cycle_transfer
python collect_data.py --data_type 'base' --data_id 1

2.Modify the xml file to change the agent morphology.

cd cross_morphology/base_train_test/mujoco_xml
python modify_xml.py --env 'half_cheetah' --phy '3leg'
cd ../../cycle_transfer
python collect_data.py --data_type '3leg' --data_id 1

3.Train the model to learn the correspondence.

cd cross_morphology/cycle_transfer
python alignexp.py --data_type1 'base' --data_id1 1 --data_type2 '3leg' --data_id2 1

Reproducing Real Xarm Robot Experiment

Download the dataset here.

To train and evaluate the model:

cd realxarm_exp/xarm_pose
python main.py

Toy Experiment

The code for toy experiment is shown in the path './toys', which shows how we explore step by step to get the final framework. This part is not cleaned up but you can still have a reference if you are interested in.