Unity ML-Agents

ML-Agents is Unity's open-source toolkit for training game agents. You turn a Unity scene into a learning environment and define agents with observations, actions and rewards. Then you train them with reinforcement learning or imitation learning, using PyTorch in a companion Python package. Trained behavior runs back inside the scene. Unity lists automated testing of game builds as a use, alongside NPC behavior and design checks. The package ships with 17+ example environments. Agents can learn from your own recorded play, and trained models run inside builds through Unity's inference engine. Where it falls down: it's free, but you build the bots yourself. You write the agent code, set up Python and tune training. We haven't trained an agent for this page; it uses the GitHub README and Unity's docs.

A Unity developer comfortable with Python who wants free bots that replay levels and catch broken paths.

Skip it if you want bots that work out of the box; ML-Agents is free, but you build the bots yourself.

Free and open source (Apache 2.0). Needs Unity; Unity Personal is free for individuals and small organizations under $200K revenue. Checked Oct 14, 2026 on the official pricing page.

• Costs nothing. The toolkit is open source, and Unity Personal is free for small teams.
• Learns from you. Imitation learning trains agents from your recorded demonstrations.
• Python access. Gym and PettingZoo wrappers let you plug in other training code.
• Maintained. Release 23 (Sep 2, 2025), and the repo was updated this week.

• You build the bots. There's no ready-made tester; you code the agents and rewards.
• Setup is technical. It needs Unity 6, Python 3.10.12 in Conda, and matching package versions.
• Unity only. It doesn't work with Unreal, Godot or other engines.
• Training takes time. Agents must learn before they can test, and results depend on your reward design.

nunu.ai runs ready-made AI agents that play your game on real devices from a plain-language test. It has a free Starter plan with 2,000 one-time credits. ML-Agents is free forever, but you write and train the agents. nunu.ai works across engines and platforms. ML-Agents only works in Unity. ML-Agents also needs you to write the pass and fail checks yourself. Pick nunu.ai to start testing this week. Pick ML-Agents if you have the skills and want no ongoing cost.

1. Install Unity 6000.0 or later through Unity Hub.
2. Create a Conda environment with Python 3.10.12.
3. In Unity, add the ML-Agents package from the Package Manager.
4. Install the matching Python package: pip install mlagents==1.1.0.
5. Open an example environment, then train it from the terminal.
6. Add an agent to your own scene and reward it for reaching test goals.