Time Discretization-Invariant Safe Action Repetition for Policy Gradient Methods

This repository is the official implementation of

• Seohong Park, Jaekyeom Kim, Gunhee Kim. Time Discretization-Invariant Safe Action Repetition for Policy Gradient Methods. In NeurIPS, 2021.

It contains the implementations for SAR, FiGAR-C and base policy gradient algorithms (PPO, TRPO and A2C).

The code is based on Stable Baselines3 (SB3) for PPO and A2C, and Stable Baselines (SB) for TRPO.

Requirements

• Python 3.7.8
• MuJoCo 1.5

Run examples

PPO and A2C (based on Stable Baselines3)

To install requirements:

cd sb3
pip install -r requirements.txt
pip install -e .

Train SAR-PPO on InvertedPendulum-v2 with δ = 0.01:

python repeat/main.py --env=InvertedPendulum-v2 --algo=ppo_rg --frame_skip=1 --dt=0.01 --max_t=0.05 --max_d=0.5

Train FiGAR-C-PPO on InvertedPendulum-v2 with δ = 0.01:

python repeat/main.py --env=InvertedPendulum-v2 --algo=ppo_rg --frame_skip=1 --dt=0.01 --max_t=0.05

Train PPO on InvertedPendulum-v2 with the original δ:

python repeat/main.py --env=InvertedPendulum-v2 --algo=ppo_rg

Train SAR-A2C on InvertedPendulum-v2 with δ = 0.01:

python repeat/main.py --env=InvertedPendulum-v2 --algo=a2c_rg --frame_skip=1 --dt=0.01 --max_t=0.05 --max_d=0.5

Train SAR-PPO on InvertedPendulum-v2 with δ = 0.002 and the "Action Noise" setting:

python repeat/main.py --env=InvertedPendulum-v2 --algo=ppo_rg --frame_skip=1 --dt=0.002 --max_t=0.05 --max_d=0.5 --anoise_type=action --anoise_prob=0.05 --anoise_std=3

Train SAR-PPO on InvertedPendulum-v2 with δ = 0.002 and the "External Force" setting:

python repeat/main.py --env=InvertedPendulum-v2 --algo=ppo_rg --frame_skip=1 --dt=0.002 --max_t=0.05 --max_d=0.5 --anoise_type=ext_f --anoise_prob=0.05 --anoise_std=300

Train SAR-PPO on InvertedPendulum-v2 with δ = 0.002 and the "Strong External Force (Perceptible)" setting:

python repeat/main.py --env=InvertedPendulum-v2 --algo=ppo_rg --frame_skip=1 --dt=0.002 --max_t=0.05 --max_d=0.5 --anoise_type=ext_fpc --anoise_prob=0.05 --anoise_std=1000

TRPO (based on Stable Baselines)

To install requirements:

cd sb
pip install -r requirements.txt
pip install -e .

Train SAR-TRPO on InvertedPendulum-v2 with δ = 0.01:

python repeat/main.py --env=InvertedPendulum-v2 --frame_skip=1 --dt=0.01 --max_t=0.05 --max_d=0.5

License

This codebase is licensed under the MIT License. See also sb3/LICENSE_SB3 and sb/LICENSE_SB.