Requirements

pytorch==1.6.0, mujoco-py==2.0.2.13 All the requirments are specified in requirements.txt

Data Collection

Run the following script in a bash.

for seed in {1..40}
do
	python train_data_collection.py --env-type cheetah_vel --save-models 1 --log-tensorboard 1 --seed $seed
done

Data collection program uses SAC algorithm to collect offline data, hyperparameters can be accessed and modified at data_collection_config

Mujoco Environment Installation

Ant and Cheetah environments need mujoco210. Refer to https://github.com/openai/mujoco-py for more details about mujoco210 installation.

Walker and Hopper environments need mujoco131. Download mjpro131 and mjkey from https://www.roboti.us/download.html, extract them into ~/.mujoco/mjpro131, and set export MUJOCO_PY_MJPRO_PATH=~/.mujoco/mjpro131, then mujoco131 is ready to go.

Be aware that the environment variable of mujoco131 MUJOCO_PY_MJPRO_PATH is different from mujoco210 MUJOCO_PY_MUJOCO_PATH. Please discern them to avoid potential errors.

Run FOCAL

python train_offline_FOCAL.py --env-type cheeta_vel

Change the argument --env-type to choose a different environment:

Environment	Argument
Half-Cheetah-Vel	cheetah_vel
Half-Cheetah-Dir	cheetah_dir
Ant-Dir	ant_dir

Hyperparameters of FOCAL can be modified at offline_rl_config