Ce Hao, Catherine Weaver, Chen Tang, Kenta Kawamoto, Masayoshi Tomizuka, Wei Zhan
- python 3.8+
- mujoco 2.0 (for RL experiments)
- Ubuntu 18.04+
Create a virtual environment (e.g. conda) with Python>=3.8 and install the following requirements
# download the repo
git clone https://github.com/CeHao1/skill-critic.git
cd skill-critic
# install requirements and package
pip3 install -r requirements.txt
pip3 install -e .
To manage the data and checkpoints, we recommand to put them in the current directory as,
mkdir ./experiments
mkdir ./data
export EXP_DIR=./experiments
export DATA_DIR=./data
For Maze experiments, please download the demonstration data from SPiRL, at drive.
Then place the them in the ./data/point_maze
.
For Fetch robot experiments, please download demonstration data from ReSkill, at drive. Then use converter.py to convert the data format and finally put then in the ./data/reskill_fetch_robot
.
All results will be written to WandB. Before running any of the commands below, create an account and then change the WandB entity and project name at the top of train_skill.py and train_rl.py to match your account.
The programs for train Maze and Fetch robot environments are listed in the scripts. Please run the whole scripts or run individual script.
sh src/scripts/skill/point_maze.sh
sh src/scripts/skill/reskill_fetch_robot.sh
When the program finishhed, please copy the checkpoint to the weights directory. Details are in the scripts.
sh src/scripts/rl/point_maze.sh
sh src/scripts/rl/reskill_fetch_robot.sh
sh src/scripts/hrl/point_maze.sh
sh src/scripts/hrl/reskill_fetch_robot.sh
Our implementation consults some functions in the official repo of SPiRL