Adaptive Intervention Mechanism (AIM)

ICML 2025

Official release for the code used in the paper: Robot-Gated Interactive Imitation Learning with Adaptive Intervention Mechanism

Webpage | Code | Paper

Installation

git clone https://github.com/metadriverse/AIM.git
cd AIM

# Create Conda environment
conda create -n aim python=3.7
conda activate aim

# Install dependencies
pip install -r requirements.txt
pip install -e .

Launch Experiments

MetaDrive

In MetaDrive, we use a well-trained PPO neural expert to approximate human policies generated by the script train_ppo_metadrive.py. We provide a pretrained PPO expert in aim/experiments/metadrive/metadrive_ppo_20m_steps.zip.

To reproduce the main experiment reported in the paper, run the training script train_aim_metadrive.py in the folder aim/experiments/metadrive. The main experiment of AIM takes about 10 minutes. We also provide a simpler toy environment with --toy_env.

You can also specify an intended switch-to-human frequency SWITCH_RATE and initial warm-up steps WARMUP_STEPS, depending on your budget of expert involvement. The default SWITCH_RATE is 0.05, and the default WARMUP_STEPS is 200.

cd ~/AIM

# Train PPO expert (Skip this if you want)
python aim/experiments/metadrive/train_ppo_metadrive.py \
--exp_name ppo_metadrive

# Run toy experiment
python aim/experiments/metadrive/train_aim_metadrive.py \
--exp_name aim_metadrive_toy \
--toy_env

# Run full experiment
python aim/experiments/metadrive/train_aim_metadrive.py \
--exp_name aim_metadrive \
--wandb \
--wandb_project WADNB_PROJECT_NAME \
--wandb_team WANDB_ENTITY_NAME \
--delta SWITCH_RATE \
--init_bc_steps WARMUP_STEPS

Minigrid

In Minigrid, we provide a pretrained neural expert in aim/experiments/minigrid/best_model_minigrid_4roomlarge.zip to approximate human policies. The training script train_aim_minigrid.py trains AIM in the Minigrid multi-room environment.

cd ~/AIM
python aim/experiments/minigrid/train_aim_minigrid.py \
--exp_name aim_minigrid \
--wandb \
--wandb_project WADNB_PROJECT_NAME \
--wandb_team WANDB_ENTITY_NAME

References

@article{cai2025robot,
  title={Robot-Gated Interactive Imitation Learning with Adaptive Intervention Mechanism},
  author={Cai, Haoyuan and Peng, Zhenghao and Zhou, Bolei},
  journal={International Conference on Machine Learning},
  year={2025}
}