This is the implementation of the Stubborn Agent (yuumi_the_magic_cat on the Habitat Challenge Leaderboard). It uses relatively simple strategies but achieved a strong result (0.237 success rate and 0.098 SPL). We release the code here to help the Robotic community by providing them with a strong baseline that they can both compete against or work from.

The Stubborn Agent is modified from Goal-Oriented-Semantic-Policy and makes improvement in exploration strategy, untrapping strategy, and object detection strategy, which are discueed in details in the paper.

The core of the code in located in Stubborn/agent.

Stubborn/agent/stubborn_agent.py contains the implementation of the stubborn agent;

Stubborn/agent/agent_state.py contains implementation of the global goal module and goal detection module;

Stubborn/agent/mapping_module.py contains implementation of the mapping module;

Stubborn/agent/agent_helper.py contains implementation of the path planning module.

We use Docker to run the code, therefore users don't need to manually install any dependencies. Users need to download pretrained weights and environment dataset before they can run the code.

Users can download pretrained weights for RedNet here. Download the rednet_semmap_mp3d_tuned.pth file and place it in Stubborn.

Users can download data used to train the Object Detection Module here. Place obj_id_data.pickle into the Stubborn folder.

• Download the Matterport 3D dataset using the instructions here

• Move the Matterport 3D scene dataset or create a symlink at data/scene_datasets/mp3d.

• Download the Matterport 3D Object Navigation Task Data here.

• Move the task dataset to data/datasets

• The data folder should look like this:

Stubborn/
  data/
    scene_datasets/
      mp3d/
        17DRP5sb8fy
        1LXtFkjw3qL
        ...
    datasets/
      objectnav/
        mp3d/
          v1/
            train/
            val/
            val_mini/

To verify that the data is setup correctly, run:

sh remote_submission_run.sh

The agent should get 0.233 Success, 0.139 SPL and 4.301 DTG.

• This project builds on the Goal-Oriented Semantic Policy paper.
• Rednet Object Segmentation Model is trained by Joel Ye et. al at https://joel99.github.io/objectnav/.