/AnimalAI-Olympics

Code repository for the Animal AI Olympics competition

Primary LanguagePython

Animal-AI Olympics

July 1st - November 1st

The Animal-AI Olympics is an AI competition with tests inspired by animal cognition. Participants are given a small environment with just seven different classes of objects that can be placed inside. In each test, the agent needs to retrieve the food in the environment, but to do so there are obstacles to overcome, ramps to climb, boxes to push, and areas that must be avoided. The real challenge is that we don't provide the tests in advance. It's up to you to explore the possibilities with the environment and build interesting configurations that can help create an agent that understands how the environment's physics work and the affordances that it has. The final submission should be an agent capable of robust food retrieval behaviour similar to that of many kinds of animals. We know the animals can pass these tests, it's time to see if AI can too.

Prizes $32,000 (equivalent value)

  • Overall Prizes
    • 1st place overall: $7,500 total value - $6,500 with up to $1,000 travel to speak at NeurIPS 2019.
    • 2nd place overall: $6,000 total value - $5,000 with up to $1,000 travel to speak at NeurIPS 2019.
    • 3rd place overall: $1,500.
  • WBA-Prize: $5,000 total value - $4,000 with up to $1,000 travel to speak at NeurIPS 2019
  • Category Prizes: $200 For best performance in each category (cannot combine with other prizes - max 1 per team).
  • Mid-way AWS Research Credits: The top 20 entries as of September 1st will be awarded $500 of AWS credits.

See competition launch page and official rules for further details.

Important Please check the competition rules here. To submit to the competition and be considered for prizes you must also fill in this form. Entry to the competition (via EvalAI) constitutes agreement with all competition rules.

Overview

Here you will find all the code needed to compete in this new challenge. This repo contains the training environment (v1.0) that will be used for the competition. Information for entering can be found in the submission documentation. Please check back during the competition for minor bug-fixes and updates, but as of v1.0 the major features and contents are set in place.

For more information on the competition itself and to stay updated with any developments, head to the Competition Website and follow @MacroPhilosophy and @BenBeyret on twitter.

The environment contains an agent enclosed in a fixed sized arena. Objects can spawn in this arena, including positive and negative rewards (green, yellow and red spheres) that the agent must obtain (or avoid). All of the hidden tests that will appear in the competition are made using the objects in the training environment. We have provided some sample environment configurations that should be useful for training (see examples/configs), but part of the challenge is to experiment and design new configurations.

To get started install the requirements below, and then follow the Quick Start Guide. More in depth documentation can be found on the Documentation Page.

Evaluation

The competition has 300 tests, split over ten categories. The categories range from the very simple (e.g. food retrieval, preferences, and basic obstacles) to the more complex (e.g. spatial reasoning, internal models, object permanence, and causal reasoning). We have included example config files for the first seven categories. Note that the example config files are just simple examples to be used as a guide. An agent that solves even all of these perfectly may still not be able to solve all the tests in the category, but it would be off to a good start.

The submission website allows you to submit an agent that will be run on all 300 tests and it returns the overall score (number of tests passed) and score per category. We cannot offer infinite compute, so instances will be timed out after ~90 minutes and only tests performed up to that point counted (all others will be considered failed). See the submission documentation for more information.

For the mid-way and final evaluation we will (resources permitting) run more extensive testing with 3 variations per test (so 900 tests total). The variations will include minor perturbations to the configurations. The agent will have to pass all 3 variations to pass each individual test, giving a total score out of 300. This means that your final test score might be lower than the score achieved during the competition and that the competition leaderboard on EvalAI may not exactly match the final results.

Development Blog

You can read the launch posts - with information about prizes and the categories in the competition here:

Animal-AI Evaluation: July 8th - with collated information about the evaluation.

Animal-AI Launch: July 1st - with information about the prizes and introduction to all 10 categories.

You can read the development blog here. It covers further details about the competition as well as part of the development process.

  1. Why Animal-AI?

  2. The Syllabus (Part 1)

  3. The Syllabus (Part 2): Lights Out

Requirements

The Animal-AI package works on Linux, Mac and Windows, as well as most Cloud providers. Note that for submission to the competition we only support linux-based Docker files.

We recommend using a virtual environment specifically for the competition. You will need python3.6 installed (we currently only support python3.6). Clone this repository to run the examples we provide.

We offer two packages for this competition:

  • The main package is an API for interfacing with the Unity environment. It contains both a gym environment as well as an extension of Unity's ml-agents environments. You can install it via pip:

    pip install animalai

    Or you can install it from the source, head to animalai/ folder and run pip install -e .

  • We also provide a package that can be used as a starting point for training, and which is required to run most of the example scripts found in the examples/ folder. At the moment we only support Linux and Max for the training examples. It contains an extension of ml-agents' training environment that relies on OpenAI's PPO, as well as Google's dopamine which implements Rainbow (among others). You can also install this package using pip:

    pip install animalai-train

    Or you can install it from source, head to examples/animalai_train and run pip install -e .

Finally download the environment for your system:

OS Environment link
Linux download v1.0.0
MacOS download v1.0.0
Windows download v1.0.0

You can now unzip the content of the archive to the env folder and you're ready to go! Make sure the executable AnimalAI.* is in env/. On linux you may have to make the file executable by running chmod +x env/AnimalAI.x86_64. Head over to Quick Start Guide for a quick overview of how the environment works.

Manual Control

If you launch the environment directly from the executable or through the VisualizeArena script it will launch in player mode. Here you can control the agent with the following:

Keyboard Key Action
W move agent forwards
S move agent backwards
A turn agent left
D turn agent right
C switch camera
R reset environment

Citing

Official Animal-AI Papers Coming Soon. In the meantime please cite the Nature: Machine Intelligence piece for any work involving the competition environment.

Crosby, M., Beyret, B., Halina M. The Animal-AI Olympics Nature Machine Intelligence 1 (5) p257 2019.

Unity ML-Agents

The Animal-AI Olympics was built using Unity's ML-Agents Toolkit.

The Python library located in animalai is almost identical to ml-agents v0.7. We only added the possibility to change the configuration of arenas between episodes. The documentation for ML-Agents can be found here.

Juliani, A., Berges, V., Vckay, E., Gao, Y., Henry, H., Mattar, M., Lange, D. (2018). Unity: A General Platform for Intelligent Agents. arXiv preprint arXiv:1809.02627

EvalAI

The competition is kindly hosted on EvalAI, an open source web application for AI competitions. Special thanks to Rishabh Jain for his help in settting this up.

Deshraj Yadav, Rishabh Jain, Harsh Agrawal, Prithvijit Chattopadhyay, Taranjeet Singh, Akash Jain, Shiv Baran Singh, Stefan Lee and Dhruv Batra (2019) EvalAI: Towards Better Evaluation Systems for AI Agents

Known Issues

In play mode pressing R or C does nothing sometimes. This is due to the fact that we have synchronized these features with the agent's frames in order to have frames in line with the configuration files for elements such as blackouts. Solution: press the key again, several times if needed.

When a lot of objects are spawned randomly, extremely rarely, the agent will fall through the floor. (fixed in v0.6.1)

TODO

  • Add custom resolutions
  • Add inference viewer to the environment
  • Offer a gym wrapper for training
  • Improve the way the agent spawns
  • Add lights out configurations.
  • Improve environment framerates
  • Add moving food

Version History

  • v1.0.3

    • Adds inference mode to Gym environment
    • Adds seed to Gym Environment
    • Submission example folder containing a trained agent
    • Provide submission details for the competition
    • Documentation for training on AWS

  • v1.0.2

    • Adds custom resolution for docker training as well
    • Fix version checker

  • v1.0.0

    • Adds custom resolution to both Unity and Gym environments
    • Adds inference mode to the environment to visualize trained agents
    • Prizes announced
    • More details about the competition

  • v0.6.1 (Environment only)

    • Fix rare events of agent falling through the floor or objects flying in the air when resetting an arena

  • v0.6.0

    • Adds score in playmode (current and previous scores)
    • Playmode now incorporates lights off directly (in examples try: python visualizeArena.py configs/lightsOff.yaml)
    • To simplify the environment several unnecessary objects have been removed see here
    • Several object properties have been changed also here
    • Frames per action reduced from 5 to 3 (i.e.: for each action you send we repeat it for a certain number of frames to ensure smooth physics)
    • Add versions compatibility check between the environment and API
    • Remove step_number argument from animalai.environment.step

  • v0.5 Package animalai, gym compatible, dopamine example, bug fixes

    • Separate environment API and training API in Python
    • Release both as animalai and animalai-train PyPI packages (for pip installs)
    • Agent speed in play-mode constant across various platforms
    • Provide Gym environment
    • Add trainBaselines,py to train using dopamine and the Gym wrapper
    • Create the agent.py interface for agents submission
    • Add the HotZone object (equivalent to the red zone but without death)

  • v0.4 - Lights off moved to Unity, colors configurations, proportional goals, bugs fixes

    • The light is now directly switched on/off within Unity, configuration files stay the same
    • Blackouts now work with infinite episodes (t=0)
    • The rand_colors configurations have been removed and the user can now pass RGB values, see here
    • Rewards for goals are now proportional to their size (except for the DeathZone), see here
    • The agent is now a ball rather than a cube
    • Increased safety for spawning the agent to avoid infinite loops
    • Bugs fixes

  • v0.3 - Lights off, remove Beams and add cylinder

    • We added the possibility to switch the lights off at given intervals, see here
    • visualizeLightsOff.py displays an example of lights off, from the agent's point of view
    • Beams objects have been removed
    • A Cylinder object has been added (similar behaviour to the Woodlog)
    • The immovable Cylinder tunnel has been renamed CylinderTunnel
    • UnityEnvironment.reset() parameter config renamed to arenas_configurations_input

  • v0.2 - New moving food rewards, improved Unity performance and bug fixes

    • Moving rewards have been added, two for each type of reward, see the details here.
    • Added details for the maze generator.
    • Environment performance improved.
    • Issue #7 (-inf rewards for t: 0 configuration) is fixed.

  • v0.1 - Initial Release