Spectral-Risk-Constrained Policy Optimization (SRCPO)

Requirements

• Python 3.8 or greater
• torch==1.12.1
• gymnasium
• safety_gymnasium
• scikit-learn
• pandas
• qpsolvers=1.9.0
• ruamel.yaml

How to train

python main.py --task_cfg_path tasks/{task_name}.yaml --algo_cfg_path algos/{algo_name}/{env_name}.yaml --seed {seed_idx} --gpu_idx {gpu_idx} --wandb

or

bash scripts/train.sh

How to eval

python main.py --task_cfg_path tasks/{task_name}.yaml --algo_cfg_path algos/{algo_name}/{env_name}.yaml --seed {seed_idx} --gpu_idx {gpu_idx} --model_num {model_num} --eval

or

bash scripts/eval.sh

How to test (render)

python main.py --task_cfg_path tasks/{task_name}.yaml --algo_cfg_path algos/{algo_name}/{env_name}.yaml --seed {seed_idx} --gpu_idx {gpu_idx} --model_num {model_num} --test