Learning to Extrapolate: A Transductive Approach

Code for Bilinear Transdcution on regression and imitation learning tasks as proposed in Learning to Extrapolate: A Transductive Approach

Setup

Clone the repository

git clone https://github.com/avivne/bilinear-transduction.git

Create a virtual environment with the following requirements

conda env create -f environment.yml

Download the grasping data and place demos.pkl under data/grasping.

Training and Evaluation

Training and evaluating bilinear transduction on regression tasks (grasping)

python bilinear_transduction_regression.py

Training and evaluating bilinear transduction on imitation learning tasks

python bilinear_transduction_imitation.py --config-name configs/reach_metaworld.yaml

replace reach_metaworld with push_metaworld, slider or adroit for other imitation learning environments.

Add --model-type bc for training and evaluating the neural net baseline on any of the tasks.

Render Videos

Rendering sample videos of expert demonstrations and model performance on in-distribution and OOS data points

python render_videos.py

modify this block to render selected environments, models and data points.

Cite

If you use this code in your research, please consider citing

@inproceedings{netanyahu2023transduction,
  title={Learning to Extrapolate: A Transductive Approach},
  author={Netanyahu, Aviv and Gupta, Abhishek and Simchowitz, Max and Zhang, Kaiqing and Agrawal, Pulkit},
  booktitle={The Eleventh International Conference on Learning Representations (ICLR)},
  year={2023}
}

Acknowledgements

The environments and data are derived from Meta-World, Adroit relocate, mjrl and NDF.