This is the code for our CoRL 2020 paper PixL2R: Guiding Reinforcement Learning Using Natural Language by Mapping Pixels to Rewards

1. Create and activate conda environment using the environment.yml file:

conda env create -f environment.yml
conda activate pixl2r

2. Install dependencies for metaworld, by running the following command in metaworld/ directory:

pip install -e .

3. Download and extract preprocessed data, by running the following commands in PixL2R directory:

wget https://www.cs.utexas.edu/users/ml/PixL2R/data.zip
unzip data.zip -d data/

(The downloaded file is ~3GB; the extracted files require 50GB of disk space.)

1. Supervised learning of PixL2R model: Run the following command in src/supervised

python model.py --save-path=<save-path>

2. Policy training without language-based rewards: Run the following command in src/rl

python train_policy.py --obj-id=6 --env-id=1 --reward-type=<sparse|dense>

3. Policy training with language-based rewards: Pass the PixL2R model filepath (output of step 1) and the description id to use.

python train_policy.py --obj-id=6 --env-id=1 --reward-type=<sparse|dense> --model-file=/path/to/PixL2R/model --descr-id=<0|1|2>

The raw videos can be downloaded from the following link: https://www.cs.utexas.edu/users/ml/PixL2R/videos.zip
The videos were generated using the script src/rl/generate_videos.py.

The codebase is based on the following repositories:

1. https://github.com/rlworkgroup/metaworld
2. https://github.com/nikhilbarhate99/PPO-PyTorch