Repolode
Submit
Unity-Technologies

Unity-Technologies/ml-agents

Open-source toolkit from Unity that lets games and simulations serve as environments for training AI agents with reinforcement and imitation learning.

View on GitHubWebsite
Stars
19,769
Forks
4,512
Open issues
19
Last commit
6 days ago

What is it?

ML-Agents is a toolkit from Unity Technologies that turns Unity scenes into training environments for intelligent agents. Developers add the Unity SDK to a game or custom scene, define agent observations, actions and rewards, and train behavior with a Python layer built on PyTorch.

The toolkit includes deep reinforcement learning algorithms such as PPO, SAC and MA-POCA (for multi-agent cooperative and competitive scenarios, plus self-play), and imitation learning through behavioral cloning and GAIL. It also supports curriculum learning, environment randomization, and running multiple Unity instances concurrently. Trained models can be run inside Unity through its Inference Engine.

Python users can control Unity environments directly or wrap them as Gym or PettingZoo environments. The project ships with 17+ example environments and is aimed at game developers building NPC behavior or automated tests, as well as researchers evaluating AI methods in Unity.

Key features

  • 01Supports PPO, SAC, MA-POCA and self-play for single and multi-agent training.
  • 02Provides behavioral cloning and GAIL for learning from demonstrations.
  • 03Includes curriculum learning and environment randomization.
  • 04Runs trained models natively in Unity via the Inference Engine.
  • 05Exposes Unity environments as Gym and PettingZoo environments.
  • 06Ships with 17+ example Unity environments.

Use cases

  • Training NPC behavior for adversarial or multi-agent game scenarios.
  • Automated testing of game builds using trained agents.
  • Evaluating game design decisions before release through simulated play.
  • Researching reinforcement learning methods in a rich 3D or VR simulation platform.

Topics

#reinforcement-learning#imitation-learning#unity#game-ai#multi-agent#pytorch#simulation#deep-learning#deep-reinforcement-learning#machine-learning

FAQ

What is ml-agents?
Open-source toolkit from Unity that lets games and simulations serve as environments for training AI agents with reinforcement and imitation learning. ML-Agents is a toolkit from Unity Technologies that turns Unity scenes into training environments for intelligent agents. Developers add the Unity SDK to a game or custom scene, define agent observations, actions and rewards, and train behavior with a Python layer built on PyTorch.
Is ml-agents free to use? What license does it have?
ml-agents is open source under the Other license. Check the license terms before using it in commercial products.
Is ml-agents still maintained?
Actively maintained. The last commit to ml-agents was on Oct 13, 2026.
What can I use ml-agents for?
Training NPC behavior for adversarial or multi-agent game scenarios. Automated testing of game builds using trained agents. Evaluating game design decisions before release through simulated play. Researching reinforcement learning methods in a rich 3D or VR simulation platform.
Where can I get ml-agents?
The source code and installation instructions are on GitHub: https://github.com/Unity-Technologies/ml-agents

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects