Abstract
The Animal-AI Environment is a novel testbed for artificial intelligence inspired by the field of comparative psychology. Rather than chasing high scores in specific video games, this platform evaluates an agent's "common sense" and ability to generalize across 300 tasks designed to probe cognitive milestones like object permanence, causal reasoning, and spatial memory. By dissociating training from testing and focusing on functional affordances, the environment challenges agents to develop robust internal models of the physical world. Our results demonstrate that while curriculum learning improves generalization, a significant gap remains between current AI capabilities and human-level reasoning.
Introduction
While Deep Reinforcement Learning (DRL) has achieved superhuman performance in complex games like Go and StarCraft, these successes often mask a lack of true understanding. Agents typically excel only within the specific distributions they were trained on; minor changes to the environment layout often lead to performance collapse. They lack the "animal-like" ability to adapt to unexpected situations or reason about the physical world.
The authors argue that current AI benchmarks are often task-oriented (focused on winning a specific game) rather than ability-oriented (testing fundamental understanding, such as object permanence). To bridge this gap, they introduce an environment that forces agents to develop functional cognitive traits—such as intuitive physics and planning—that are transferable across diverse scenarios.
Methods: The Animal-AI Arena
Built on the Unity ML-Agents platform, the environment features a simple, consistent physical setup designed to test cognitive affordances rather than visual pattern matching.
The environment consists of:
- The Agent: A blue sphere with monocular pixel vision and a basic sense of its own velocity.
- The Arena: A 40x40 area where various objects can be spawned.
- Objects: Immovable obstacles (walls, ramps), rewards (green/gold spheres for "food"), negative stimuli (red spheres/zones), and movable objects (boxes and shapes).
- Configurability: Researchers can precisely define or randomize the placement, size, and color of objects using YAML configuration files, creating a versatile "laboratory" for testing.
The methodology emphasizes curriculum learning, where an agent is trained on a progression of increasingly complex physical layouts to encourage the development of generalizable skills rather than rote memorization of a single map.
The Animal-AI Testbed
The benchmark consists of 300 hidden test configurations categorized into 10 cognitive domains derived from animal cognition literature. These include basic food retrieval, preferences, obstacle navigation, avoidance, spatial reasoning, robustness, internal models (memory), object permanence, numerosity, and causal reasoning. By using tests that animals like corvids or even nematodes can pass, the researchers provide a clear ladder of cognitive complexity for AI to climb.
Key Results and Generalization
A key insight of this work is the dissociation of training from testing. Unlike standard RL benchmarks, the Animal-AI test battery is kept secret, requiring agents to learn the underlying physics and logic of the world during training to succeed on novel puzzles.
Baseline experiments compared an agent trained on a specific 2x2 maze against one trained on a curriculum of varying layouts:
- The Overfitting Trap: The agent trained only on the 2x2 maze achieved 99% success on its training task but plummeted to 41% when faced with a circular maze.
- The Power of Curriculum: The curriculum-trained agent was significantly more robust, maintaining consistent performance across different maze types, such as scrambled layouts.
- The AI Gap: Top entries in the "Animal-AI Olympics" competition scored roughly 40% overall, while human performance is near 100%, highlighting a massive remaining gap in "common sense" reasoning for current AI.
Conclusion
The Animal-AI Environment serves as a bridge between the AI community and comparative psychologists, providing a standardized way to measure the "intelligence" of artificial agents against the natural world. Future work will expand the environment to include complex substrates (liquids, ropes), multi-agent interaction, and online learning requirements.
Summary updated