Hierarchical GAIL

Official implemention of the paper Hierarchical Generative Adversarial Imitation Learning with Mid-level Input Generation for Autonomous Driving on Urban Environments.

Youtube Playlist hGAIL Supplementary Videos.

Installation

1. Install python3.8.
2. Install required packages from requirements.txt file.

Reproducing results

• To collect data from expert run the start_expert.sh script.

• To train the agents from the paper set the following variables global variables from the file rl_train.py.

• The variable FAKE_BIRDVIEW controls if the agent will use the birdview from the env or generated by the GAN.

• The variable GAIL controls if the agent will be trained using GAIL or PPO.

• The variable RGB_GAIL controls if the agent trains from birdview images or directly from images generated by the cameras.

• The variable RESUME_LAST_TRAIN controls if the agent will start a new training or resume the last one.

• To train an agent using GAIL or PPO run the start_rl.sh script after setting the appropriate variables.

• To train an agent using Behavior Cloning run the start_bc.sh script.

• To train only the GAN birdview using the samples from the expert demonstrations run the start_gan.sh script.

Results

• The results are logged using Wandb (Weights & Biases) that is similar to an online user based tensorboard.

• The training is logged in runs, where one csv table with data from every step is generated for each environment.

• The agent parameters are saved in a pytorch file every epoch in the ckpt directory identified by the epoch number of steps.

• To evaluate an agent run the eval_agent.sh script after setting the appropriate variables from in file evaluate_agent.py.

Generating plots

• In the paper_plots directory there are many scripts to generate the plots used in the paper

Generating videos

• To generate videos from the agent evaluation use the generate_video.py python script.

Disclaimer

This project is built upon the code from the following projects:

• Official code release of the paper
  End-to-End Urban Driving by Imitating a Reinforcement Learning Coach
  

• Official code release of the paper
  Augmenting GAIL with BC for sample efficient imitation learning
  

• PyTorch-GAN

Citation

Please consider citing our paper if you find it useful or want to use it in your research:

@misc{hGAIL,
      title={Hierarchical Generative Adversarial Imitation Learning with Mid-level Input Generation for Autonomous Driving on Urban Environments}, 
      author={Gustavo Claudio Karl Couto and Eric Aislan Antonelo},
      year={2023},
}