This repository contains the PushWorld environment, a collection of benchmark puzzles to compare the performance of different solvers, and a state-of-the-art planner for this environment.
Benchmark puzzles and reference solutions are available in the benchmark
directory. Puzzles are stored in a human-readable format that is convenient to
edit in any text editor with a fixed-width font. To interactively play any of
the puzzles in this benchmark, visit https://deepmind-pushworld.github.io/play/.
The cpp directory contains the code of the Recursive Graph Distance (RGD)
planner, which as of January 2023 is the state-of-the-art in solving the most
puzzles within 1, 5, and 30 minutes. Implementing the RGD planner in C++
ensures a fair comparison to other classical planners used in our experiments.
To build and run the RGD planner, follow the instructions in cpp/README.md.
The python3 directory contains a suite of scripts to benchmark the RGD planner,
plot videos of puzzle solutions, plot images of all puzzles, convert puzzles to
PDDL, and convert puzzles to SAS. To set up the Python environment required to
run code in the python3 directory, follow the instructions in python3/README.md.
To support reinforcement learning research with PushWorld, it also implements
interfaces compatible with the OpenAI Gym environment API and
DeepMind RL environment API. To see how to
use these, see python3/scripts/demo_gym_env.py and
python3/scripts/demo_dm_env.py, respectively.
-
PDDL: A predicate logic language for representing classical planning tasks.
-
SAS: A state-variable representation for classical planning tasks.
-
RGD: Recursive graph distance - A search heuristic with optimized distance computations.
-
OpenAI Gym: An environment for developing and testing learning agents.
-
DeepMind env: An environment for developing and testing learning agents.