Auxiliary Tasks Speed Up Learning PointGoal Navigation

Our code is built off of the Habitat framework and only changes some files. For a framework overview, please refer to Habitat documentation. This code was last tested with this commit.

Requirements

1. Python 3.6+
2. Recommended: anaconda/miniconda

Installation

Please first initialize the conda environment using the provided environment.yml file with

conda env create -f environment.yml.

To install Habitat, follow installation guidelines for habitat-sim and habitat-api. No additional packages are used in our paper. This codebase is known to be compatible with https://github.com/facebookresearch/habitat-sim/releases/tag/v0.1.6 (we use headless installation).

Data

We use the Gibson dataset as distributed by Habitat. You'll need to download the Gibson scene data from habitat-sim and the Gibson PointNav data from habitat-api. Download and extract according to the instructions on the Habitat readmes.

Pre-trained weights

NOTICE: Note that the reported results have changed since the initial pre-print in Summer 2020. An early stopped checkpoint for models trained to 40 million frames for the baseline, best single module (reported as "Add" in the paper) (cpca-id-td_single) and multi-module (cpca-id-td_attn-e) checkpoints are posted here. Note that our reported scores were attained by running ./scripts/eval_pn.sh 3 times and averaging the scores. Once your checkpoints are downloaded, modify the path in eval_pn.sh to these paths to get validation metrics. We get the following numbers:

• baseline: .56 SPL.
• cpca-id-td_single : .68 SPL.
• cpca-id-td_attn-e: .71 SPL.

Training

You can train the same models using ./scripts/train_pn.sh "variant_name. The variant configurations are all under habitat_baselines/config/official_gc. The datasets used by these configurations are specified by the task configuration under configs/tasks/pointnav_gc.yaml.

Evaluation

You can evaluate a trained checkpoint by configuring the checkpoint path in ./scripts/eval_pn.sh. To generate videos of your agent, enable VIDEO_OPTION in the variant configuration. You can also find settings to adjust what videos are generated inside ppo_trainer.py.

Model Analysis

To generate a json with more detailed evaluation information (e.g. actions taken, agent location per episode), see run_detailed.py.

References and Citation

If you use this work, you can cite it as

@inproceedings{ye2020auxiliary,
    title={Auxiliary Tasks Speed Up Learning PointGoal Navigation},
    author={Joel Ye and Dhruv Batra and Erik Wijmans and Abhishek Das},
    year={2020},
    booktitle={Proceedings of the Conference on Robot Learning (CoRL)}
}