Author's official implementation of TPAMI paper "Generalizable Multi-modal Adversarial Imitation Learning for Non-stationary Dynamics"

1. Code of the discriminator is in the file GMAIL/algps/gaifo.py, code of the generator is in the directory GMAIL/algos/escp, and code of the non-stationary environment is in the directory GMAIL/envs.

2. Get demonstrations with the following command. The expert trajectories will be stored in the directory expert_data/sac_HalfCheetah_gravity if you set the environment name as HalfCheetah and the changing parameter as gravity.

   python generate_expert_data.py

3. Run GMAIL in HalfCheetah with gravity as the changing parameter with the following command. Demonstrations are in the directory expert_data/sac_HalfCheetah_gravity, so please collect demonstrations before run the code.

   python run_gmail.py --env_name HalfCheetah-v3 --varying_params gravity --expert-path-dir expert_data/sac_HalfCheetah_gravity --H_step 4 --use_rmdm --stop_pg_for_ep --bottle_neck --rbf_radius 3000 --name_suffix GMAIL --rnn_fix_length 16 --autoalpha

4. Modify the parameters env_name and varying_params to run GMAIL in other tasks and with other varying params. For example, run GMAIL in Hopper with body mass as the changing parameter with the following command. Here, demonstrations are in the directory expert_data/sac_Hopper_mass.

   python run_gmail.py --env_name Hopper-v3 --varying_params body_mass --expert-path-dir expert_data/sac_Hopper_mass --use_absorbing_state --H_step 4 --use_rmdm --stop_pg_for_ep --bottle_neck --rbf_radius 3000 --name_suffix GMAIL --rnn_fix_length 2