ajshah91 / ltl2action

This is the code repository accompanying the ICML 2021 paper LTL2Action: Generalizing LTL Instructions for Multi-Task RL (https://arxiv.org/abs/2102.06858).

LTL2Action

This is the code repository accompanying the paper LTL2Action: Generalizing LTL Instructions for Multi-Task RL, which will appear in ICML 2021.

In the animation below, our agent's behaviour, compared to a myopic one is demonstrated on ZoneEnv, a custom environment based on OpenAI's SafetyGym:

Installation instructions

We recommend using Python 3.6 to run this code.

1. pip install -r requirements.txt
2. Install Spot-2.9
   • Follow the installation instructions at the link. Spot should be installed in /usr/local/lib/python3.6/site-packages/spot. This step usually takes around 20 mins.
3. (Optional) To run the OpenAI Safety Gym, you will need Mujoco installed, as well as an active license.
   • pip install mujoco-py==2.0.2.9
   • pip install -e src/envs/safety/safety-gym/

Training Agents

Instructions for training and evaluating RL agents on each of our domains is available in the src folder.

Citation

@article{DBLP:journals/corr/abs-2102-06858,
  author    = {Pashootan Vaezipoor and
               Andrew C. Li and
               Rodrigo Toro Icarte and
               Sheila A. McIlraith},
  title     = {LTL2Action: Generalizing {LTL} Instructions for Multi-Task {RL}},
  journal   = {CoRR},
  volume    = {abs/2102.06858},
  year      = {2021},
  url       = {https://arxiv.org/abs/2102.06858},
  archivePrefix = {arXiv},
  eprint    = {2102.06858},
  timestamp = {Thu, 18 Feb 2021 15:26:00 +0100},
  biburl    = {https://dblp.org/rec/journals/corr/abs-2102-06858.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}