This contains code for running the experiments in Policy Architectures for Compositional Generalization in Control.
Download the weights for generating BC data from this link. Unzip the file in ./weights
.
Then generate behavior cloning data by running:
./script/make_data.sh
Sweep over different environments and architectures.
python launch_bc.py -m +bc_experiment=big_transformer,deepset,mlp +bc_setup=3p,3s,2s2p
In general, setup
specifies the environment and exploration schedule and experiment
specifies the architecture. Some examples:
# 3 push for transformer and MLP (padded for extrapolation eval)
python launch.py -m +experiment=3pdense_fastexp +setup=transformer,padded_mlp seed="range(5)"
# 3 push for deepset uses a faster exploration schedule
python launch.py -m +experiment=3pdense_fastexp +setup=deepset seed="range(5)"
The majority of the code for "Entity Factored RL" is licensed under CC-BY-NC, however portions of the project are available under separate license terms.
- Portions based on Hindsight Experience Replay, Stable Baselines 3, and Gym are licensed under the MIT license.