Learning Robust Policy against Disturbance in Transition Dynamics via State-Conservative Policy Optimization

This is the code of paper Learning Robust Policy against Disturbance in Transition Dynamics via State-Conservative Policy Optimization. [arXiv]

Instructions

• install requirements (python=3.6):

pip install -r requirements.txt

• run sc-sac in the Walker2d-v2 task with default configs:

python launch.py -e Walker2d-v2 

• run sac in the HalfCheetah-v2 task:

python launch.py -n sac -e HalfCheetah-v2 

• run pr-sac in the Hopper-v2 task:

python launch.py -n pr-sac -e Hopper-v2

• plot heatmap of sc-sac trained policies in the HalfCheetah-v2 task:

python plot.py HalfCheetah-v2 sc-sac /path/to/data/save/dir

• note that before ploting the heatmap, you have to manually replace the codes in /path_to_gym_module/envs/mujoco with the codes in ./mujoco_env_enhancing_codes, which enables us to change the relative mass and the relative friction of these environments during test.

cp -r ./mujoco_env_enhancing_codes/* /path_to_gym_module/envs/mujoco/