This repository houses the code for "CAMPs: Learning Context-Specific Abstractions for Efficient Planning in Factored MDPs" by Rohan Chitnis*, Tom Silver*, Beomjoon Kim, Leslie Pack Kaelbling, and Tomás Lozano-Pérez. Conference on Robot Learning (CoRL) 2020.
Paper: https://arxiv.org/abs/2007.13202
Video: https://www.youtube.com/watch?v=wTXt6djcAd4
Installation of dependencies:
- First, run
pip install -r requirements.txt - Next, you will need to install the FF classical planner (https://fai.cs.uni-saarland.de/hoffmann/ff.html) and set the environment variable FF_PATH to point to the
ffexecutable. If you are on Linux, this can be accomplished via:- curl https://fai.cs.uni-saarland.de/hoffmann/ff/FF-v2.3.tgz --output FF-v2.3.tgz
- tar -xzvf FF-v2.3.tgz && cd FF-v2.3 && make && cd ..
- export FF_PATH="FF-v2.3/ff"
To run, simply execute python main.py. This will, by default, start training in the TAMP NAMO environment. For other environments, planners, and/or methods, change family_to_run, solver_name, and/or approach_names respectively in settings.py.
Learned context-specific independences are packaged with this code, in the learned_csi/ directory. The core CAMP abstraction is implemented in approaches/modelbased.py, with supporting code in csi.py.