/LearningToLearn

Collection of algorithms to learn loss and reward functions via gradient-based bi-level optimization.

Primary LanguageJupyter NotebookMIT LicenseMIT

LearningToLearn

This repository contains code for

  • ML3: Meta-Learning via Learned Losses, presented at ICPR 2020, won best student award (pdf)
  • MBIRL: Model-Based Inverse Reinforcement Learning from Visual Demonstrations, presented at CoRL 2020 (pdf)

Setup

In the LearningToLearn folder, run:

conda create -n l2l python=3.7
conda activate l2l 
python setup.py develop

ML3 paper experiments and citation

To reproduce results of the ML3 paper follow the README instructions in the ml3 folder

Citation

@inproceedings{ml3,
author    = {Sarah Bechtle and Artem Molchanov and Yevgen Chebotar and Edward Grefenstette and Ludovic Righetti and Gaurav Sukhatme and Franziska Meier},
title     = {Meta Learning via Learned Loss},
booktitle = {International Conference on Pattern Recognition, {ICPR}, Italy, January 10-15, 2021},
year      = {2021} }

MBIRL paper experiments and citation

To test our MBIRL algorithm follow the README instructions in the mbirl folder

Citation

@InProceedings{mbirl,
  author    = {Neha Das, Sarah Bechtle, Todor Davchev, Dinesh Jayaraman, Akshara Rai and Franziska Meier},
  booktitle = {Conference on Robot Learning (CoRL)},
  title     = {Model Based Inverse Reinforcement Learning from Visual Demonstration},
  year      = {2020},
  video     = {https://www.youtube.com/watch?v=sRrNhtLk12M&t=52s},
}

License

LearningToLearn is released under the MIT license. See LICENSE for additional details about it. See also our Terms of Use and Privacy Policy.