Awesome

A dataset for Image-Goal Navigation in Habitat

This repo contains the training & evaluation data used in the paper

Memory-Augmented Reinforcement Learning for Image-Goal Navigation [Paper]

Description

The train (resp. evaluation) set is split in 72 (resp. 14) files corresponding to the scenes in Habitat. Each scene contains episodes of various difficulty (Easy, Medium Hard & Extra) depending on the start/goal distance. The histogram of number of trajectories with respect to start-goal distance per split is shown below.

traj_distrib_plot

Citation

If you use this data in your research, please consider citing the paper as follows

@article{mezghani2021memory,
  title={Memory-augmented reinforcement learning for image-goal navigation},
  author={Mezghani, Lina and Sukhbaatar, Sainbayar and Lavril, Thibaut and Maksymets, Oleksandr and Batra, Dhruv and Bojanowski, Piotr and Alahari, Karteek},
  journal={arXiv preprint arXiv:2101.05181},
  year={2021}
}

License

image-goal-nav-dataset is CC-BY-NC 4.0 licensed, as found in the LICENSE file.

Some information in the data may be derived from the Gibson dataset, which is available under the following license.