Your Unity scenes are now reinforcement learning gyms

What it does ML-Agents is Unity’s official open-source toolkit for training intelligent agents inside games and simulations. It bakes PyTorch-based implementations of algorithms like PPO, SAC, and MA-POCA into a Unity SDK, letting developers train NPCs that can adapt, automate playtesting, or prototype game design decisions without leaving the engine. Researchers get a Python API and wrappers for Gym and PettingZoo, so the same Unity scenes can feed into standard RL pipelines.

The interesting bit The toolkit treats your game loop as the environment and your NPCs as the agents, collapsing the usual wall between game development and AI research. It supports everything from solo agents to adversarial multi-agent brawls, and even lets you swap in your own Python training algorithms if the built-in trainers feel too prescriptive.

Key highlights

• Ships with 17+ example environments and supports 2D, 3D, and VR/AR scenes
• Covers single-agent, cooperative, and competitive multi-agent setups out of the box
• Includes both deep RL (PPO, SAC, MA-POCA, self-play) and imitation learning (BC, GAIL)
• Offers curriculum learning and environment randomization for training under varied conditions
• Native cross-platform inference engine and support for concurrent environment instances
• Verified Unity package for Unity 2020.1 and later

Caveats

• Primary documentation has migrated to Unity Package docs; the old web docs are deprecated, so update your bookmarks
• The develop branch is explicitly flagged as unstable
• In-editor analytics are enabled to “improve developer experience,” which may matter if you’re privacy-sensitive

Verdict A strong fit for Unity developers who need trainable NPCs, automated testing, or rapid AI prototyping; pure ML researchers without Unity scenes to spare can likely find lighter-weight alternatives.