This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.

---

Implementation of PVR for Control, as presented in The (Un)Surprising Effectiveness of Pre-Trained Vision Models for Control.

The main branch reproduces the results presented in the paper.
The distributed branch uses multiprocessing and distributed training for much faster learning. It also has wrappers for other environments, including Adroit and DMC. Results are slightly different, but we encourage to use it if you would like to build upon our paper.

Codebase Installation

conda create -n pvr_habitat python=3.8
conda activate pvr_habitat
git clone git@github.com:sparisi/pvr_habitat.git
cd pvr_habitat
pip install -r requirements.txt

Habitat Installation

• Clone https://github.com/sparisi/habitat-lab and do a full install with habitat_baselines.

The main differences between this and the original Habitat repository are:

1. STOP action removed,
2. Bugfix where the agent is placed slightly above the ground, and therefore the terminal goal condition is never triggered.

• Download and extract Replica scenes in the root folder of pvr_habitat.

WARNING! The dataset is very large!

sudo apt-get install pigz
git clone https://github.com/facebookresearch/Replica-Dataset.git
cd Replica-Dataset
./download.sh replica-path

If you have already downloaded it somewhere else, just make a symbolic link

ln -s path/to/Replica-Dataset Replica-Dataset

How to Run Experiments

There are three main scripts to run behavioral cloning:

• main_bc_1.py loads raw trajectories saved as pickles, passes observations (images) through the embedding, and then learns on the embedded observations.
• main_bc_2.py directly loads embedded observations that have already been passed through the embedding, in order to save time.
• main_bc_finetune.py is used to finetune the random PVR.

For more details on how to generate trajectories and pickles, see the README in the behavioral_cloning folder.

Pre-trained models can be downloaded here.