Hierarchical Adversarial Inverse Reinforcement Learning

Codebase for my paper: Hierarchical Adversarial Inverse Reinforcement Learning

Language: Python

The following parts are included:

Benchmarks built with Mujoco, including Hopper, Walker, Ant box-pushing, and Point maze.
An implementation of the hierarchical imitation learning (HIL) algorithm proposed in our paper.
Implementations of the SOTA IL and HIL algorithms as baselines, including GAIL, AIRL, Option-GAIL, Directed-Info GAIL.

The paper is available at: https://arxiv.org/abs/2210.01969

Please cite this paper:

@article{DBLP:journals/corr/abs-2210-01969,
  author       = {Jiayu Chen and
                  Tian Lan and
                  Vaneet Aggarwal},
  title        = {Hierarchical Adversarial Inverse Reinforcement Learning},
  journal      = {CoRR},
  volume       = {abs/2210.01969},
  year         = {2022},
  url          = {https://doi.org/10.48550/arXiv.2210.01969},
  doi          = {10.48550/arXiv.2210.01969}
}

How to config the environments:

on Ubuntu 18.04
python 3.6
pytorch 1.6
tensorboard 2.5
mujoco_py >= 1.5
gym == 0.19.0
matplotlib
tqdm
seaborn
...

Experiments with Hopper

You need to first enter the folder 'HierAIRL_Hopper'.
To run the code with specific algorithms:

# Option-GAIL:
python ./run_baselines.py --env_type mujoco --env_name Hopper-v2 --n_demo 1000 --device "cuda:0" --tag option-gail-1k --algo option_gail

# GAIL:
python ./run_baselines.py --env_type mujoco --env_name Hopper-v2 --n_demo 1000 --device "cuda:0" --tag gail-1k --algo gail

# DI-GAIL:
python ./run_baselines.py --env_type mujoco --env_name Hopper-v2 --n_pretrain_epoch 50 --n_demo 1000 --device "cuda:0" --tag d_info_gail-1k --algo DI_gail

# Option-AIRL
python ./run_main.py --env_type mujoco --env_name Hopper-v2 --n_demo 1000 --device "cuda:0" --tag option-airl-1k --algo option_airl

# H-AIRL
python ./run_main.py --env_type mujoco --env_name Hopper-v2 --n_demo 1000 --device "cuda:0" --tag hier-airl-1k --algo hier_airl

# H-GAIL
python ./run_main.py --env_type mujoco --env_name Hopper-v2 --n_demo 1000 --device "cuda:0" --tag hier-gail-1k --algo hier_gail

To run the code with the random seed Y, for which we simply choose 0, 1, or 2, please add '--seed=Y' to the back. The same below for other tasks.
For the hyperparameters, please refer to 'HierAIRL_Hopper/default_config.py'. The same below for other tasks.

Experiments with Walker

You need to first enter the folder 'HierAIRL_Walker'.
To run the code with specific algorithms:

# Option-GAIL:
python ./run_baselines.py --env_type mujoco --env_name Walker2d-v2 --n_demo 5000 --device "cuda:0" --tag option-gail-5k --algo option_gail

# GAIL:
python ./run_baselines.py --env_type mujoco --env_name Walker2d-v2 --n_demo 5000 --device "cuda:0" --tag gail-5k --algo gail

# DI-GAIL:
python ./run_baselines.py --env_type mujoco --env_name Walker2d-v2 --n_pretrain_epoch 50 --n_demo 5000 --device "cuda:0" --tag d_info_gail-5k --algo DI_gail

# Option-AIRL
python ./run_main.py --env_type mujoco --env_name Walker2d-v2 --n_demo 5000 --device "cuda:0" --tag option-airl-5k --algo option_airl

# H-AIRL:
python ./run_main.py --env_type mujoco --env_name Walker2d-v2 --n_demo 5000 --device "cuda:0" --tag hier-airl-5k --algo hier_airl

# H-GAIL:
python ./run_main.py --env_type mujoco --env_name Walker2d-v2 --n_demo 5000 --device "cuda:0" --tag hier-gail-5k --algo hier_gail

Experiments with AntPusher

You need to first enter the folder 'HierAIRL_Ant'.
To run the code with specific algorithms:

# Option-GAIL:
python ./run_baselines.py --env_type mujoco --env_name AntPusher-v0 --n_demo 10000 --device "cuda:0" --tag option-gail-10k --algo option_gail

# GAIL:
python ./run_baselines.py --env_type mujoco --env_name AntPusher-v0 --n_demo 10000 --device "cuda:0" --tag gail-10k --algo gail

# DI-GAIL:
python ./run_baselines.py --env_type mujoco --env_name AntPusher-v0 --n_pretrain_epoch 100 --n_demo 10000 --device "cuda:0" --tag d_info_gail-10k --algo DI_gail

# Option-AIRL:
python ./run_main.py --env_type mujoco --env_name AntPusher-v0 --n_demo 10000 --device "cuda:0" --tag option-airl-10k --algo option_airl

# H-AIRL:
python ./run_main.py --env_type mujoco --env_name AntPusher-v0 --n_demo 10000 --device "cuda:0" --tag hier-airl-10k --algo hier_airl

# H-GAIL:
python ./run_main.py --env_type mujoco --env_name AntPusher-v0 --n_demo 10000 --device "cuda:0" --tag hier-gail-10k --algo hier_gail

Experiments on Point Room/Corridor

You need to first enter the folder 'HierAIRL_Point'.
To reproduce the results of expert trajectories, please run the following command, where XXX can be Point4Rooms-v1 or PointCorridor-v1. The results will be available in the folder 'result'.

python ./plot_options_exp.py --env_type mujoco --env_name XXX

To reproduce the results of trajectories of the learned agents, please run the following command, where XXX can be Point4Rooms-v1 or PointCorridor-v1. The results will be available in the folder 'result'.

python ./plot_options.py --env_type mujoco --env_name XXX

To reproduce the learned agents with H-AIRL (i.e., the checkpoints), please run the following command, where XXX can be Point4Rooms-v1 or PointCorridor-v1. The results will be available in the folder 'result'.

python ./run_main.py --env_type mujoco --env_name XXX --n_demo 5000--device "cuda:0" --tag hier-airl-5k --algo hier_airl

Transfer Learning Results on Point Room

You need to first enter the folder 'HierAIRL_Point_Room_transfer'.
To run the code with specific algorithms, please run the following commands, where X can be 0, 1, 2.

# Option-GAIL:
python ./run_baselines.py --env_type mujoco --env_name Point4Rooms-v1 --n_demo 5000 --device "cuda:0" --tag option-gail-5k --algo option_gail --seed X

# GAIL:
python ./run_baselines.py --env_type mujoco --env_name Point4Rooms-v1 --n_demo 5000 --device "cuda:0" --tag gail-5k --algo gail --seed X

# H-AIRL:
python ./run_main.py --env_type mujoco --env_name Point4Rooms-v1 --n_demo 5000 --device "cuda:0" --tag hier-airl-5k --algo hier_airl --seed X --init 0

# H-AIRL initialized with the checkpoint trained in another task:
python ./run_main.py --env_type mujoco --env_name Point4Rooms-v1 --n_demo 5000 --device "cuda:0" --tag hier-airl-5k --algo hier_airl --seed X --init 1

Transfer Learning Results on Point Corridor

You need to first enter the folder 'HierAIRL_Point_Corridor_transfer'.
To run the code with specific algorithms, please run the following commands.

# Option-GAIL:
python ./run_baselines.py --env_type mujoco --env_name PointCorridor-v1 --n_demo 5000 --device "cuda:0" --tag option-gail-5k --algo option_gail --seed X

# GAIL:
python ./run_baselines.py --env_type mujoco --env_name PointCorridor-v1 --n_demo 5000 --device "cuda:0" --tag gail-5k --algo gail --seed X

# H-AIRL:
python ./run_main.py --env_type mujoco --env_name PointCorridor-v1 --n_demo 5000 --device "cuda:0" --tag hier-airl-5k --algo hier_airl --seed X --init 0

# H-AIRL initialized with the checkpoint trained in another task:
python ./run_main.py --env_type mujoco --env_name PointCorridor-v1 --n_demo 5000 --device "cuda:0" --tag hier-airl-5k --algo hier_airl --seed X --init 1

LucasCJYSDL/HierAIRL

Hierarchical Adversarial Inverse Reinforcement Learning

The paper is available at: https://arxiv.org/abs/2210.01969

How to config the environments:

Experiments with Hopper

Experiments with Walker

Experiments with AntPusher

Experiments on Point Room/Corridor

Transfer Learning Results on Point Room

Transfer Learning Results on Point Corridor