IL-CGL: Imitation Learning Agents guided by a Coverage-based Gaze Loss (CGL)

This respository contains code for the following paper:

A. Saran, R. Zhang, E.S. Short, and S. Niekum. Efficiently Guiding Imitation Learning Agents with Human Gaze. International Conference on Autonomous Agents and Multiagent Systems (AAMAS), May 2021.

Data

• Download the demonstrations and corresponding gaze data from the Atari-HEAD dataset. Store the data for each game under a separate subdirectory inside the data/atari-head/ folder.
• Download the pretrained gaze models (trained on all demos except the high-score ones for each game) and store them inside the data/trained_gaze_models/ folder.
• Note that you can re-train human gaze prediction models using the network architecture provided in BC-CGL/human_gaze.py if needed.

Code

Please see further instructions to train gaze-augmented models for:

• Behavioral Cloning in the BC-CGL folder
• Trajectory-ranked reward extrapolation in the TREX-CGL folder

Bibliography

If you find this repository is useful in your research, please cite the paper:

@article{saran2020efficiently,
  title={Efficiently Guiding Imitation Learning Agents with Human Gaze},
  author={Saran, Akanksha and Zhang, Ruohan and Short, Elaine Schaertl and Niekum, Scott},
  booktitle={International Conference on Autonomous Agents and Multiagent Systems (AAMAS)},
  year={2021},
  organization={IFAAMAS}
}