Instructions coming soon.
This branch contains the experiments for the Online Action Recognition paper, published in AAAI 2021.
This project uses pddlgym as a submodule. To clone, use
git clone --recurse-submodules https://github.com/sprkrd/sat_strips_learn.git
If you've already cloned the repository and you want to fetch the submodules too, use
git submodule update --init --recursive
More info here.
In order for PDDLGym to work with planner support, you must set the
FF_PATH env variable to the path of the Fast Forward executable.
You can also add the following to your ~/.bashrc
export FF_PATH=/path/to/ff/executable
We recommend setting up a virtual environment for installing the dependencies:
virtualenv sat_strips_learn_venv -p python3
source sat_strips_learn_venv/bin/activate
pip install -r requirements.txt
Run the following to see a small demo of OARU in the sokoban domain (the
same example used in Figure 1 of the paper):
demo_with_pddlgym.py
This will create an out folder, with several transxxx subfolder showing
how each transition is processed. Each subfolder contains: (1) a coarse
view of the Action Unification tree; (2) a detailed view of the same tree;
(3) two images representing the states before and after the transition.
Also, in out an animated GIF showing the whole execution can be found.
To view the characteristics of the benchmark domains and problems (Tables 1, 2 and 3 from the supplement), please run:
./benchmark_environments.py
The tables will be directly printed on the terminal. This shouldn't take much time (around 10s).
To run the performance and accuracy experiments (Tables 1 and 2 and Figure 2 from the paper), and obtain the action libraries presented in the supplement, run:
./benchmark_performance_and_accuracy.py
The two tables and the plots can be found in the newly created out folder, as
tabular_full_obs.tex, tabular_partial_obs.tex, full_obs_updates.pdf,
and partial_obs_updates. The action libraries sit under the partial_observability
and full_observability folders, also in out.
The plot from Figure 2, in the supplement, may be generated running:
./benchmark_observation_sampling.py
Two figures will be generated in the out folder:
observation_sampling_partial_obs_plot.pdf and
observation_sampling_full_obs_plot.pdf.
Our implementation contains many additional features that, for lack of space,
could not be detailed in the paper. The most important ones are two: (1) an Action
Unification broadphase to discard clearly non-unifiable actions before resorting
to the WPMS machinery; and (2) a labeled predicate filter to discard clearly
unrelated predicates in a TGA and ease the Action Unification process. The
source code for Action Unification and OARU can be found inside the
satstripslearn package.