Abstract
Recent advances in artificial intelligence have been strongly driven by the use of game environments for training and evaluating agents. Games are often accessible and versatile, with well-defined state-transitions and goals allowing for intensive training and experimentation. However, agents trained in a particular environment are usually tested on the same or slightly varied distributions, and solutions do not necessarily imply any understanding. If we want AI systems that can model and understand their environment, we need environments that explicitly test for this. Inspired by the extensive literature on animal cognition, we present an environment that keeps all the positive elements of standard gaming environments, but is explicitly designed for the testing of animal-like artificial cognition.
Introduction
Current artificial intelligence (AI) excels at mastering specific, complex games like Go or StarCraft II, often reaching superhuman performance. However, these agents are notoriously brittle, struggling to generalize to even slightly different environments. If an agent is trained and tested on the same distribution of tasks, it may simply be memorizing a complex policy rather than developing a true "understanding" of the world.
The authors argue that current benchmarks lack environments that explicitly test for cognitive capacities—the fundamental building blocks of intelligence that allow humans and animals to adapt to unexpected situations. To bridge this gap, they look to the field of comparative psychology, which has spent over a century refining controlled experiments to probe the mental lives of animals without the "noise" of complex controls or high-fidelity graphics.
The Animal-AI Environment
The authors developed the Animal-AI Environment using Unity ML-Agents. This physics-based sandbox is designed to test whether AI agents possess "animal-like" cognitive abilities such as object permanence and causal reasoning.
- The Agent: A simple sphere with monocular (first-person) pixel input and a basic sense of its own speed. It can move forward/backward and rotate.
- The Arena: A 3D space governed by a deterministic physics engine (gravity, friction, collisions).
- The Objects: Seven categories of items, including immovable walls and ramps, movable boxes and "tools" (U-shaped and L-shaped objects), and reward objects (green/gold spheres for "food," red spheres/zones for "poison").
Unlike standard benchmarks where training and testing are nearly identical, this environment encourages training on a "playground" of affordances—learning how to push, climb, or avoid—and then testing on entirely unseen, secret configurations.
The Animal-AI Testbed
The paper introduces a battery of 300 hidden tests used in the "Animal-AI Olympics." These tests are categorized into 10 cognitive domains:
- Basic food retrieval: Finding food in an open space.
- Preferences: Choosing the larger or more accessible reward.
- Obstacles: Navigating around walls.
- Avoidance: Staying away from "poison" zones.
- Spatial Reasoning: Understanding the environment's layout.
- Robustness: Solving problems when visual textures change.
- Internal Models: Navigating when the lights are turned off.
- Object Permanence: Knowing an object still exists when it's behind a wall.
- Numerosity: Distinguishing between different quantities of rewards.
- Causal Reasoning: Planning several steps ahead to achieve a goal.
Key Insights: Ability-Oriented Evaluation
A core insight of this work is that AI evaluation should shift from task-oriented to ability-oriented metrics, mirroring animal testing. We should not merely care if an agent can solve one specific maze; we care if it understands the concept of a maze.
The authors demonstrate the curriculum advantage: an agent trained on a fixed 2x2 maze fails when the maze becomes circular or scrambled. However, an agent trained on a curriculum—a series of increasingly complex, randomized wall configurations—develops much better generalization, even if it never encountered a "circular" maze during training.
Furthermore, AI systems do not come pre-built with an understanding of intuitive physics. While a crow might instinctively know how to use a tool, an AI must be explicitly trained in a "playground" to understand that objects can be pushed or used to bridge gaps.
Results and Performance Gap
The benchmark reveals a massive gap between current reinforcement learning successes and human-level general intelligence:
- Human vs. AI: Humans score nearly 100% across all 300 tests. Animals like chimps and corvids (crows) can solve the most complex causal reasoning tasks.
- Current AI State: At the time of the paper's "Olympics" half-way point, the top AI entries were scoring around 40%.
- Domain Variance: AI agents performed relatively well on navigation and basic reward-seeking but struggled significantly with Object Permanence and Causal Reasoning.
- Generalization: In head-to-head maze tests, curriculum-trained agents achieved near parity with specialized agents on simple tasks but significantly outperformed them on "scrambled" and "circular" mazes where specialized performance plummeted.
Limitations and Future Work
The current environment is a first iteration. Future work aims to address several limitations:
- Physical Complexity: The environment lacks liquids, soft bodies, and ropes—elements common in advanced animal cognition studies.
- Social Intelligence: The current setup is single-agent. Future versions aim to include multi-agent scenarios to test for social learning and cooperation.
- Online Learning: Most current agents are "frozen" after training. True animal-like cognition might require agents that can learn and adapt during the test trial itself.
Summary updated