spriteworld

Spriteworld is a python-based reinforcement learning environment consisting of a 2-dimensional arena with objects that can be freely moved.


Keywords
ai, reinforcement-learning, python, machine, learning, objects, deepmind, environment, generative, procedural, rl, sprites
License
Apache-2.0
Install
pip install spriteworld==1.0.2

Documentation

Spriteworld: A Flexible, Configurable Reinforcement Learning Environment

Description

Spriteworld is a python-based RL environment that consists of a 2-dimensional arena with simple shapes that can be moved freely. This environment was developed for the COBRA agent introduced in the paper "COBRA: Data-Efficient Model-Based RL through Unsupervised Object Discovery and Curiosity-Driven Exploration" (Watters et al., 2019). The motivation for the environment was to provide as much flexibility for procedurally generating multi-object scenes while retaining as simple an interface as possible.

Spriteworld sprites come in a variety of shapes and can vary continuously in position, size, color, angle, and velocity. The environment has occlusion but no physics, so by default sprites pass beneath each other but do not collide or interact in any way. Interactions may be introduced through the action space, which can update all sprites each timestep. For example, the DiscreteEmbodied action space (see spriteworld/action_spaces.py) implements a rudimentary form of physics in which an agent's body sprite can adhere to and carry sprites underneath it.

There are a variety of action spaces, some of which are continuous (like a touch-screen) and others of which are discrete (like an embodied agent that takes discrete steps).

Example Tasks

Below are three of the tasks used in the COBRA paper.

Goal-finding task. The agent must bring the target sprites (squares) to the center of the arena.

goal_finding_video

Clustering task. The agent must arrange the sprites into clusters according to their color.

clustering_video

Sorting task. The agent must sort the sprites into goal locations according to their color (each color is associated with a different goal location).

sorting_video

Installation

Spriteworld can be installed using pip:

pip install spriteworld

Or through Github:

pip install git+https://github.com/deepmind/spriteworld.git

or alternatively by checking out a local copy of our repository and running:

git clone https://github.com/deepmind/spriteworld.git
pip install spriteworld/

This last option additionally downloads tests, the demo UI and an example run loop.

Getting Started

Prerequisites

Spriteworld depends on numpy, six, absl, PIL, matplotlib, sklearn, and dm_env.

Running The Demo

Once installed, you may familiarize yourself with Spriteworld through run_demo.py:

python /path/to/local/spriteworld/run_demo.py

This will run a cluster-by-color task with a drag-and-drop action space. There are a number of tasks specified in the spriteworld/configs directory, each of which can be run with the demo by modifying the --config flag. Note that some tasks (namely spriteworld.configs.examples.goal_finding_embodied) use an embodied agent instead of the drag-and-drop action space.

Creating Your Own Task

In spriteworld/tasks.py are three tasks: FindGoalPosition, Clustering, and MetaAggregated. These can be configured and combined in numerous ways to create a wide variety of tasks, including all of those used in the COBRA paper. In particular, see spriteworld/configs/cobra/sorting.py for a non-trivial combination of goal-finding tasks.

You may create new tasks be re-using these building blocks, or creating entirely new types of tasks (just be sure to inherit from spriteworld/tasks.AbstractTask).

Running An Agent

See example_run_loop.py for an example of how to run a random agent on a Spriteworld task. See spriteworld/gym_wrapper.py if you prefer the OpenAI Gym environment interface.

Additional Use-Cases

Spriteworld can be used for purposes other than reinforcement learning. For example, it was used to generate the image datasets with controlled factor distributions presented in the paper "Spatial Broadcast Decoder: A Simple Architecture for Learning Disentangled Representations in VAEs" (Watters et al., 2019). It can also be easily extended to generate datasets of objects interacting with simple physical forces (e.g. spring, gravity, etc.), which are useful for research in unsupervised learning of visual dynamics.

Reference

If you use this library in your work, please cite it as follows:

@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",
}

Disclaimers

This is not an officially supported Google product.