Spriteworld-Physics implements some basic physical forces and dynamics on top of Spriteworld. It can be used to generate simple videos of objects moving and interacting.
Spriteworld is a python-based RL environment consisting of a 2-dimensional arena with simple shapes that can be moved around. See [https://github.com/deepmind/spriteworld] for a more detailed description.
This library ignores the RL/agent aspect of Spriteworld, only using it for video generation. This library introduces a number of forces/dynamics into the environment, namely springs, gravity/magents, and collisions. See forces.py for details.
Below are examples of the kinds of systems Spriteworld-Physics can simulate. See
configs/ for example configs and run_demo.py for how to visualize them.
Springs. Four sprites fully connected by springs. See configs/springs.py .
Collisions. A random number of sprites that collide and bounce off walls. See
configs/collisions.py .
Magnets. Four circles of various masses/charges that magnetically repel each
other and bounce off walls. See configs/magnets.py .
Star system. Stars orbiting a fixed center circle according to gravity. This
demonstrates the use of a non-fully-connected interaction graph. See
configs/star_system.py .
Colliding springs. Circles connected by springs that also collide with each
other. This demonstrates the use of two force laws simultaneously. See
configs/colliding_springs.py .
Spriteworld-Physics can be installed through Github:
pip install git+https://github.com/nwatters01/spriteworld-physics.gitor alternatively by checking out a local copy of our repository and running:
git clone https://github.com/nwatters01/spriteworld-physics.git
pip install spriteworld-physics/This last option additionally downloads the demo UI and an example gif generator script.
Spriteworld-Physics depends on spriteworld , numpy , six , absl , and
dm_env .
Once installed, you may familiarize yourself with Spriteworld-Physics through
run_demo.py :
python /path/to/local/spriteworld-physics/run_demo.pyThis will run a system of objects interacting with springs and play a video of
the environment. There are a number of other example systems defined by configs
in spriteworld_physics/configs/ . Run the demo with flag
--config=$path_to_task_config$ to view each of these.
All available forces lie in spriteworld_physics/forces.py and various simple
strategies for defining interaction graphs (i.e.specifying which forces should
apply between which objects) lie in spriteworld_physics/graph_generators.py .
You can configure your own system by combining these, or if necessary implement
your own force or graph generator classes.
See the examples in spriteworld_physics/configs/ for demonstrations of the
current version of this codebase's scope.
There is currently no pipeline for large-scale data-generation. However, there
is a script spriteworld_physics/generate_gif.py which runs a config and writes
a video of the resulting simulation to a file as a gif, which you can adapt to
generate data in your preferred format.
This library is derivative work from Spriteworld:
@misc{spriteworld19,
author = {Nicholas Watters and Loic Matthey and Sebastian Borgeaud and Rishabh Kabra and Alexander Lerchner},
title = {Spriteworld: A Flexible, Configurable Reinforcement Learning Environment},
howpublished = {https://github.com/deepmind/spriteworld/},
url = {https://github.com/deepmind/spriteworld/},
year = "2019",
}
See the Spriteworld license in LICENSE-apache-2.0-derivative.
If you use this library in your work, please cite Spriteworld as above.