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

Implementation of Change-Based Exploration Transfer (C-BET), as presented in Interesting Object, Curious Agent: Learning Task-Agnostic Exploration.

This code was built on the RIDE repository.

Codebase and MiniGrid Installation

conda create -n cbet python=3.8.10
conda activate cbet
git clone git@github.com:sparisi/cbet.git
cd cbet
pip install -r requirements.txt

Habitat Installation (not Needed for MiniGrid Experiments)

• Follow the official guide and do a full install with habitat_baselines.
• Download and extract Replica scenes in the root folder of cbet

  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 the script does not work, manually unzip with cat replica_v1_0.tar.gz.part* | tar -xz

How to Run Experiments

• Intrinsic-only pre-training: OMP_NUM_THREADS=1 python main.py --model cbet --env <ENV_NAME> --no_reward --intrinsic_reward_coef=0.005

• Extrinsic-only transfer with pre-trained model: OMP_NUM_THREADS=1 python main.py --model cbet --env <ENV_NAME> --intrinsic_reward_coef=0.0 --checkpoint=path/to/model.tar
  Pre-trained exploration policies are available as code release.

• Tabula-rasa training with summed intrinsic and extrinsic reward: OMP_NUM_THREADS=1 python main.py --model cbet --env <ENV_NAME> --intrinsic_reward_coef=0.005

See src/arguments.py for the full list of hyperparameters.

For MiniGrid, <ENV_NAME> can be MiniGrid-DoorKey-8x8-v0, MiniGrid-Unlock-v0, ...
For Habitat, <ENV_NAME> can be HabitatNav-apartment_0, HabitatNav-hotel_0, ...

• Offline testing (single): python test_checkpoint.py --checkpoint=<PATH_TO_TAR>

• Offline testing (batch): python test_models.py --logdir=<PATH_TO_LOG> (this tests all models in a log directory produced by a Slurm batch run).