Adversairial Skil Learning for Robust Manipulation

Author: Pingcheng Jian, Chao Yang, Di Guo, Huaping Liu, Fuchun Sun

We provide the code for Adversairial Skil Learning for Robust Manipulation in this repository.

Requirements

• This codebase requires Robel and other dependencies in your conda environment with pip:

$ pip install -e .

• if you want to install some extra development tools, using:

$ pip install -e .[dev]

Instruction to train

1. train the DoublePick-v1

python train/train.py --algo adversarial_double_pick_sac --env-name DoublePick-v1 --cuda True --seed 1

2. train the protagonist and adversary in DClawTurnFixed-v0

python train/adversarial_robel_train.py

3. train normal SAC policy in DClawTurnFixed-v0

python train/robel_train.py

Instruction to test

1. test protagonist against adversary in DoublePick-v1

python test/adversarial_test_double_pick.py --algo adversarial_double_pick_sac --env-name DoublePick-v1 --seed 121

2. test normal SAC policy against adversary in DoublePick-v1

python test/adversarial_test_double_pick_normal.py --algo adversarial_double_pick_sac --env-name DoublePick-v1 --seed 121

3. test protagonist or normal SAC policy against random attack in DoublePick-v1

python test/adversarial_test_double_pick_random_noise.py --algo adversarial_double_pick_sac --env-name DoublePick-v1 --seed 121

4. test protagonist against adversary in DClawTurnFixed-v0

python test/ad_test_robel.py --ad_factor 0.5 --env-name DClawTurnFixed-v0 --po_agent robust

5. test normal SAC policy against adversary in DClawTurnFixed-v0

python test/ad_test_robel.py --ad_factor 0.5 --env-name DClawTurnFixed-v0 --po_agent normal

6. test protagonist against random noise in DClawTurnFixed-v0

python test/ad_test_robel.py --ad_factor 0.5 --env-name DClawTurnFixed-v0 --po_agent robust

7. test normal SAC policy against random noise in DClawTurnFixed-v0

python test/ad_noise_test_robel.py --ad_factor 0.5 --env-name DClawTurnFixed-v0 --po_agent normal