ml-agents

The Unity Machine Learning Agents Toolkit (ML-Agents) is an open-source project by Unity Technologies that bridges game development and artificial intelligence research. It provides a flexible platform where games and simulations act as training environments for intelligent agents, using deep reinforcement learning, imitation learning, neuroevolution, and other methods. Key capabilities include support for single-agent, multi-agent cooperative, and multi-agent competitive scenarios through several deep reinforcement learning algorithms such as PPO, SAC, MA-POCA, and self-play. For imitation learning, it offers Behavioral Cloning (BC) and Generative Adversarial Imitation Learning (GAIL). The toolkit ships with 17+ example Unity environments and a flexible Unity SDK that can be integrated into existing games or custom scenes. The Python API allows researchers to train agents using reinforcement learning, imitation learning, or any custom method. Trained agents can be deployed for multiple purposes including controlling NPC behavior in multi-agent and adversarial settings, automated testing of game builds, and evaluating game design decisions pre-release. The toolkit also supports curriculum learning scenarios, environment randomization for robust agent training, on-demand decision making, and multiple concurrent Unity environment instances for parallel training. The ML-Agents Toolkit utilizes Unity's Inference Engine for native cross-platform support and provides APIs to wrap Unity learning environments as gym or PettingZoo environments. Custom training algorithms and components can be added through a plugin system. The project is licensed under Apache 2.0 and is verified for Unity 2020.1 and later. The latest release (Release 23, August 2025) includes Python package version 1.1.0 and Unity package version 4.0.0.