What it does Animal-AI is a Unity-based reinforcement learning environment where an agent navigates a small arena collecting green rewards, avoiding red ones, and reasoning about objects that appear, disappear, or move. The repo bundles 900 pre-built tasks inspired by decades of animal cognition literature—think delayed gratification, spatial reasoning, and object permanence, but for neural networks.

The interesting bit The twist is the source of the tests: instead of optimizing for game scores, you’re benchmarking whether an RL agent exhibits cognitive skills we already know animals have. The environment exposes both a Gym interface and Unity ML-Agents, and the task battery is configurable between episodes.

Key highlights

• 900 tasks across categories like memory, navigation, and causal reasoning
• Gym-compatible API plus ML-Agents extension (based on ml-agents v0.15.0)
• Training package wraps PPO and SAC from OpenAI/BAIR baselines
• Pre-built executables for Windows, Mac, and Linux
• Jupyter notebooks for environment walkthrough and training setup

Caveats

• Not actively maintained; v3 lives in a separate repo (animal-ai by mdcrosby)
• Linux v2.0.2 fixes agent speed bugs; other platforms stuck on v2.0.1
• Requires downloading and unzipping a platform-specific Unity executable manually

Verdict Worth a look if you’re researching animal-like AI or need a structured, non-arcade RL benchmark. Skip it if you want a maintained, plug-and-play environment—this is a 2019 competition artifact with dusty corners.