GeneralizeX

Setup

python3 -m venv env
source env/bin/activate
pip install -e .

Setup robosuite private macro file:

python $PWD/env/lib/python3.9/site-packages/robosuite/scripts/setup_macros.py

Run

Compositional RL Training

There is the option to either train the baseline or compositional reinforcement learning policy.

python scripts/train.py

Flags (see scripts\train.py for more flags)

• --env: name of gymnasium environment
• --dir: directory for saved models (When resuming training, this is the directory with your trained models).
• --model_prefix: optional prefix for saved models
• --resume_training: flag to load trained models and continue training. If this flag is not provided then by default start training a new policy.

Example: train baseline lift policy (IIWA arm)

source shell_scripts/train.sh

Evaluation

python scripts/eval.py --env=ENV_NAME --dir=PATH_TO_MODEL

Example

source shell_scripts/eval.sh

Experiments

Algorithm 1

E.g. Domain-specific generalization between reach policies on two different robot arms.

    python scripts/learn_maps.py --input_env=CompLift-IIWA --input_policy=PPO --input_dir=experiments/PPO/CompLift-IIWA/20231219-145654-id-7627/models --output_env=CompLift-Panda --output_policy=PPO --output_dir=experiments/PPO/CompLift-Panda/20231222-172458-id-1179/models --epochs=1000 --type=linear