yangrui2015/awgcsl

Code for ICLR 2022 paper Rethinking Goal-Conditioned Supervised Learning and Its Connection to Offline RL.

Weighted Goal-conditioned Supervised Learning (WGCSL)

Code for ICLR 2022 paper Rethinking Goal-Conditioned Supervised Learning and Its Connection to Offline RL (website). WGCSL is a simple but effective algorithm for offline goal-conditioned Reinforcement Learning via weighted supervised learning.

We provide offline goal-conditioned benchmark with offline dataset in the 'offline_data' folder, including 'random' and 'expert' settings. The 'buffer.pkl' is used for WGCSL and other algorithms included in our codes (GCSL, MARVIL, BC, HER, DDPG, Actionable Models), and each item in the buffer are also provided as *.npy files for training Goal BCQ and Goal CQL. Due to the storage limitation, the full offline dataset is in this anonymous google drive link: https://drive.google.com/drive/folders/1SIo3qFmMndz2DAnUpnCozP8CpG420ANb.

Requirements

python3.6+, tensorflow, gym, mujoco, mpi4py

Installation

• Clone the repo and cd into it

• Install baselines package

  pip install -e .

Usage

Environments: PointReach, PointRooms, Reacher, SawyerReach, SawyerDoor,FetchReach, FetchSlide, FetchPick, HandReach.

WGCSL:

python -m  wgcsl.run  --env=FetchReach --num_env 1 --mode supervised --su_method gamma_exp_adv_clip10_baw  --load_path ./offline_data/expert/FetchReach/  --offline_train  --load_buffer  --log_path ./${path_name}

Note: For random datasets of harder tasks (from FetchPush to HandReach), we suggest using 'exp_adv_clip10_baw' instead of 'gamma_exp_adv_clip10_baw'.

GCSL:

python -m  wgcsl.run  --env=FetchReach --num_env 1 --mode supervised --load_path ./offline_data/expert/FetchReach/  --offline_train  --load_buffer

Goal MARVIL

python -m  wgcsl.run  --env=FetchReach --num_env 1 --mode supervised  --load_path ./offline_data/random/FetchReach/ --load_buffer --offline_train  --su_method exp_adv  --no_relabel True 

Goal Behavior Cloning

python -m  wgcsl.run  --env=FetchReach --num_env 1 --mode supervised  --load_path ./offline_data/expert/FetchReach/ --load_buffer --offline_train   --no_relabel True 

offline HER

python -m  wgcsl.run  --env=FetchReach --num_env 1 --mode her  --load_path ./offline_data/expert/FetchReach/ --load_buffer --offline_train   

offline DDPG

python -m  wgcsl.run  --env=FetchReach --num_env 1 --mode her  --load_path ./offline_data/expert/FetchReach/ --load_buffer --offline_train   --no_relabel True 

Ablations

GCSL + Discount Relabeling Weight:

python -m  wgcsl.run  --env=FetchReach  --num_env 1 --load_path ./offline_data/expert/FetchReach/ --load_buffer --offline_train  --mode supervised --su_method gamma

GCSL + Goal-conditioned Exponential Advantage Weight:

python -m  wgcsl.run  --env=FetchReach --num_env 1 --mode supervised --load_path ./offline_data/expert/FetchReach/ --load_buffer --offline_train  --su_method exp_adv_clip10

GCSL + Best-advantage Weight

python -m  wgcsl.run  --env=FetchReach --num_env 1 --mode supervised --load_path ./offline_data/expert/FetchReach/ --load_buffer --offline_train  --su_method baw

Citation

If you use WGCSL in your work, please cite:

@inproceedings{
yang2022rethinking,
title={Rethinking Goal-Conditioned Supervised Learning and Its Connection to Offline {RL}},
author={Rui Yang and Yiming Lu and Wenzhe Li and Hao Sun and Meng Fang and Yali Du and Xiu Li and Lei Han and Chongjie Zhang},
booktitle={International Conference on Learning Representations},
year={2022},
url={https://openreview.net/forum?id=KJztlfGPdwW}
}