“Unity: A General Platform for Intelligent Agents,” by Arthur Juliani and colleagues at Unity Technologies (first posted to arXiv in September 2018), introduced the Unity ML-Agents Toolkit and argued that modern game engines are the right foundation for training AI. The paper proposes a taxonomy of simulation platforms and contends that “modern game engines are uniquely suited to act as general platforms” for learning environments “rich in visual, physical, task, and social complexity.”
The motivation was that many existing AI training environments suffered from “unrealistic visuals, inaccurate physics, low task complexity, restricted agent perspective, or limited capacity for interaction.” A full game engine like Unity brings high-quality rendering, a real physics engine, and flexible scene authoring, and is not a closed “black-box” - researchers and developers can configure every part of the world. The accompanying open-source ML-Agents Toolkit exposes Unity scenes to a Python API so agents can be trained with reinforcement learning, imitation learning, and other methods.
ML-Agents mattered because it lowered the barrier to building custom 3D training environments and connected the large community of game developers to machine learning. It let studios train in-game behaviors and let researchers spin up rich simulations without writing an engine from scratch, making the game engine a standard tool for embodied and agent-based AI research.