Home

Awesome

Curious Representation Learning for Embodied Intelligence

This is the pytorch code for the paper Curious Representation Learning for Embodied Intelligence. This codebase is based on the codebase from Habitat-lab, please see HABITAT_README.md for installation instructions for the repository.

Interactive Pretraining of Embodied Agents

To pretrain agent weights on Matterport3D, please use the following command:

python habitat_baselines/run.py --run-type=train --exp-config habitat_baselines/cvpr_config/pretrain/curiosity_pointnav_pretrain.yaml

The other configs used in the paper may also be found in habitat_baselines/cvpr_config/pretrain.

Downstream ImageNav Pretraining

To finetune weights on ImageNav, please use the following command:

python habitat_baselines/run.py --run-type=train --exp-config habitat_baselines/cvpr_config/imagenav/curiosity_pointnav_gibson_imagenav.yaml

Downstream ObjectNav Pretraining

To finetune weights on ObjectNav, please use the following command:

python habitat_baselines/run.py --run-type=train --exp-config habitat_baselines/cvpr_config/objectnav/curiosity_pointnav_mp3d_objectnav.yaml

Pretrained Weights

The pretrained CRL model from the Matterport3D environment can be downloaded from here

Citing Our Paper

If you find our code useful for your research, please consider citing the following paper, as well as papers included in HABITAT_README.md.

@inproceedings{du2021curious,
    author = {Du, Yilun and Gan, Chuang and
    Isola, Phillip},
    title = {Curious Representation Learning
    for Embodied Intelligence},
    booktitle = {Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
    year = {2021}
}