Xander Sun¹, Louis Law, Hoyard Zhi, Ronghe Qiu¹, and Junwei Liang¹

¹ AI Thrust, The Hong Kong University of Science and Technology (Guangzhou)

We present PSL, a zero-shot approach for learning instance level navigation skill. The agent is tasked with a language goal to find a specified object in current scene, for example, a chair made of black leather and is located near two windows.

InstanceNav v.s ObjectNav.

Specifically, a semantic enhanced PSL agent is proposed and a prioritized semantic training strategy is introduced to select goal images that exhibit clear semantic supervision and relax the reward function from strict exact view matching. At inference time, a semantic expansion inference scheme is designed to preserve the same granularity level of the goal-semantic as training.

Model Architecture for PSL.

All the required data can be downloaded from here.

1. Create a conda environment:

   conda create -n psl python=3.7 cmake=3.14.0
   
   conda activate psl
   

2. Install pytorch version 1.10.2:

   pip install torch==1.10.2+cu113 torchvision==0.11.3+cu113 -i https://download.pytorch.org/whl/cu113
   

3. Install habitat-sim:

   conda install habitat-sim-challenge-2022 headless -c conda-forge -c aihabitat
   

4. Install habitat-lab:

   git clone --branch challenge-2022 https://github.com/facebookresearch/habitat-lab.git habitat-lab-challenge-2022
   
   cd habitat-lab-challenge-2022
   
   pip install -r requirements.txt
   
   # install habitat and habitat_baselines
   python setup.py develop --all 
   
   cd ..

1. Setup steps

   pip install -r requirements.txt
   
   python setup.py develop
   

2. Follow the instructions here to set up the data/scene_datasets/ directory. gibson scenes can be found here.

3. Download the HM3D objectnav dataset from ZSON.

   wget https://dl.fbaipublicfiles.com/habitat/data/datasets/objectnav/hm3d/v1/objectnav_hm3d_v1.zip
   
   unzip objectnav_hm3d_v1.zip -d data/datasets/objectnav/
   
   # clean-up
   rm objectnav_hm3d_v1.zip

4. Download the HM3D instance navigation dataset.

   # download the original Instance Image Navigation dataset
   wget https://dl.fbaipublicfiles.com/habitat/data/datasets/imagenav/hm3d/v3/instance_imagenav_hm3d_v3.zip
   
   unzip instance_imagenav_hm3d_v3.zip -d data/datasets/
   
   # clean-up
   rm instance_imagenav_hm3d_v3.zip
   
   mkdir -p data/datasets/instancenav/val
   
   # download the attribute descriptions
   wget --no-check-certificate "https://drive.google.com/uc?export=download&id=1KNdv6isX1FDZi4KCVPiECYDxijg9cZ3L" -O data/datasets/instancenav/val/val_text.json.gz
   
   export PROJECT_ROOT=`pwd`
   cd data/datasets/instancenav/val/
   ln -s $PROJECT_ROOT/data/datasets/instance_imagenav_hm3d_v3/val/content .
   cd $PROJECT_ROOT

5. Download the trained checkpoints PSL_Instancenav.pth, and move to data/models.

6. Download the retrieved goal embeddings hm3d_objectnav.imagenav_v2.pth and hm3d_instancenav.imagenav_v2.pth

   mkdir -p data/goal_datasets/objectnav data/goal_datasets/instancenav
   
   # download retrieved goal embeddings
   wget --no-check-certificate "https://drive.google.com/uc?export=download&id=1spyyqfsSfhHL8pp5DG6aBZjozNny4IQd" -O data/goal_datasets/objectnav/hm3d_objectnav.imagenav_v2.pth
   
   wget --no-check-certificate "https://drive.google.com/uc?export=download&id=1UhA132XoQB0-4sflfBeBl-0uTt4RtidG" -O data/goal_datasets/instancenav/hm3d_instancenav.imagenav_v2.pth

1. Setup data/goal_datasets using the script tools/extract-goal-features.py. This caches CLIP goal embeddings for faster training.

   Your directory structure should now look like this:

   .
   +-- habitat-lab-challenge-2022/
   |   ...
   +-- zson/
   |   +-- data/
   |   |   +-- datasets/
   |   |   |   +-- objectnav/
   |   |   |   +-- imagenav/
   |   |   +-- scene_datasets/
   |   |   |   +-- hm3d/
   |   |   |   +-- mp3d/
   |   |   +-- goal_datasets/
   |   |   |   +-- imagenav/
   |   |   |   |   +-- hm3d/
   |   |   +-- models/
   |   |   |   +-- PSL_Instancenav.pth
   |   +-- zson/
   |   ...
   

Evaluate the PSL agent on the ObjectNav task:

bash scripts/eval/objectnav_hm3d.sh

Evaluate the PSL agent on the InstanceNav Text-Goal task:

bash scripts/eval/instancenav_text_hm3d.sh