RL Training and Evaluation

Installation

Our RL training code is based on rl-games, install the package requirements below:

pip install gymnasium ray seaborn wandb trimesh

Next, fork and local-install our custom fork of rl-games repo:

git clone https://github.com/MandiZhao/rl_games.git
cd rl_games
pip install -e . 

RL Training

Assume wandb logging is setup. Run the bash script below:

cd dexmachina
bash examples/train_rl.sh

RL Evaluation

Provide full checkpoint path saved from one of the rl training runs, e.g.

CK=logs/rl_games/inspire_hand/RUN_NAME/nn/inspire_hand.pth
python dexmachina/rl/eval_rl_games.py -B 1 --checkpoint $CK -v