[Paper] [Project Website]

Official Repo of ALDA! Zero-shot visual generalization of RL agents on various DMControl tasks using disentangled representation learning with principles of associative memory.

1. Create the conda environment and activate it

conda create -n alda python=3.8
conda activate alda

2. run or manually install the packages in dmcontrol_generalization_benchmark/setup/install_envs.sh.

cd dmcontrol_generalization_benchmark
./setup/install_envs.sh
cd ..

3. Install the packages in requirements.txt:

pip install -r requirements.txt 

In dmcontrol_generalization_benchmark/setup/config.cfg change your/path/to/ to wherever you put this repository. Follow the instructions here under the Datasets header to download the DAVIS dataset for the distracting background environment. Download and extract the dataset to alda_official/dmcontrol_generalization_benchmark/datasets/.

The specs/ folder contains yaml configs for all tasks with the default hyperparameters used for the main paper results. For example, to run ALDA on the Walker Walk task:

python -m scripts.train --experiment_spec_file specs/train_alda_walker_walk.yaml

We use Weights and Biases to record results. To enable w&b, add the --use_wandb flag and specify your w&b entity and project name with the --wandb_entity and --wandb_project flags:

python -m scripts.train --experiment_spec_file specs/train_alda_walker_walk.yaml --wandb_entity xyz --wandb_project abc

By default, the code will not let you run two experiments with the same name so that you don't accidentally overwrite an existing result. To change this behavior, add the --debug flag:

python -m scripts.train --experiment_spec_file specs/train_alda_walker_walk.yaml --debug

• This SAC implementation is based off of this repository (SVEA).

• ALDA builds directly on top of QLAE for disentanglement.