This repository provides the Animal-AI Testbed, a Unity-based environment for evaluating artificial intelligence agents on tasks inspired by animal cognition research. It offers a training environment, a training library, and 900 pre-defined tasks categorized by cognitive skills, targeting AI researchers and developers interested in comparative cognition and agent testing.

How It Works

The environment utilizes Unity ml-agents, extending version 0.15.0. Agents operate within a fixed-size arena where objects, including reward spheres (green, yellow, red), spawn. The core innovation lies in the ability to dynamically reconfigure arenas and object combinations between episodes, allowing for the creation of diverse and complex cognitive tasks. The Python API integrates with ml-agents, offering both a standard gym environment and extensions for training with PPO and SAC algorithms.

Quick Start & Requirements

• Install the API: pip install animalai
• Install the training package: pip install animalai-train
• Download the environment executable for your OS (Windows, Mac, Linux) from the provided links.
• Unzip the environment archive into the examples/env folder.
• Navigate to the examples folder and run pip install -r requirements.txt.
• Start a Jupyter notebook from the examples folder to run provided notebooks.
• Python 3 is required.

Highlighted Details

• Offers 900 tasks categorized by cognitive skills.
• Extends Unity ml-agents v0.15.0, adding arena configuration changes between episodes.
• Supports PPO and SAC training algorithms via the animalai-train package.
• Includes manual keyboard controls for direct agent interaction.

Maintenance & Community

The repository is not under active maintenance, with the last update indicating a focus on Animal-AI v3. Issues are monitored to some extent.

Licensing & Compatibility

The license is not explicitly stated in the README, but the project is associated with a competition and uses Unity ml-agents. Compatibility for commercial use or closed-source linking is not specified.

Limitations & Caveats

This codebase is for the 2019 competition and is not actively maintained. Development has shifted to a separate repository for Animal-AI v3. Only Linux has an updated environment executable (v2.0.2) addressing agent speed issues; other platforms are on v2.0.1.