HYDRO - Hybrid Cross Domain Robust Reinforcement Learning

Setup

Install requirements:

conda env create -f environment.yml

Next, you need to register the perturbed Gym environments which are placed under the folder perturbed_env. First, copy all files in the folder perturbed_env to gym/envs/mujoco. Then add the following to the file _init_.py under gym/envs:

register(
    id="HopperPerturbed-v3",
    entry_point="gym.envs.mujoco.hopper_perturbed:HopperPerturbedEnv",
    max_episode_steps=1000,
    reward_threshold=3800.0,
)

register(
    id="HalfCheetahPerturbed-v3",
    entry_point="gym.envs.mujoco.half_cheetah_perturbed:HalfCheetahPerturbedEnv",
    max_episode_steps=1000,
    reward_threshold=4800.0,
)


register(
    id="Walker2dPerturbed-v3",
    max_episode_steps=1000,
    entry_point="gym.envs.mujoco.walker2d_perturbed:Walker2dPerturbedEnv",
)

For loading offline dataset, please run the following script

python load_data.py

How to run

To run the experiments, please run the following code

python train_hydro.py --env=HalfCheetah-v3 --src_env=HalfCheetah-multi-comp --rho=0.3 --device=cuda