What is this?
Unity ML-Agents Toolkit lets Unity games and simulations act as training environments for AI agents, using deep reinforcement learning and imitation learning. It pairs a Unity SDK with a Python API so game developers and researchers can train NPCs, test builds, or evaluate design decisions.
Unity ML-Agents Toolkit lets Unity games and simulations act as training environments for AI agents, using deep reinforcement learning and imitation learning. It pairs a Unity SDK with a Python API so game developers and researchers can train NPCs, test builds, or evaluate design decisions.Is this a Unity package, a Python library, or both?Both - a flexible Unity SDK that gets integrated into your game or custom Unity scene, paired with a Python package that runs the actual RL/IL training and talks to the Unity environment over the Python API.
Both - a flexible Unity SDK that gets integrated into your game or custom Unity scene, paired with a Python package that runs the actual RL/IL training and talks to the Unity environment over the Python API.How do I get it running in my project?Add the com.unity.ml-agents Unity package (verified for Unity 2020.1+) via the Unity Package Manager, then install the matching mlagents Python package to run trainers against your scene, following the current Unity Package documentation for the latest release.
Add the com.unity.ml-agents Unity package (verified for Unity 2020.1+) via the Unity Package Manager, then install the matching mlagents Python package to run trainers against your scene, following the current Unity Package documentation for the latest release.Do I need Python and PyTorch set up separately?Yes - training runs through a separate Python package with PyTorch-based implementations of the algorithms (PPO, SAC, MA-POCA, BC, GAIL), distinct from the C# Unity-side SDK.
Yes - training runs through a separate Python package with PyTorch-based implementations of the algorithms (PPO, SAC, MA-POCA, BC, GAIL), distinct from the C# Unity-side SDK.Are the old web docs still the way to go?No - the project has deprecated the web docs; the Unity Package documentation is now the primary, up-to-date source, and the develop branch is explicitly called out as unstable.
No - the project has deprecated the web docs; the Unity Package documentation is now the primary, up-to-date source, and the develop branch is explicitly called out as unstable.Is this aimed at game developers, researchers, or both?Both - game devs get tools for NPC behavior, automated build testing, and design evaluation, while researchers get a simple Python API to try RL, imitation learning, neuroevolution, or their own custom algorithms on Unity's environments.
Both - game devs get tools for NPC behavior, automated build testing, and design evaluation, while researchers get a simple Python API to try RL, imitation learning, neuroevolution, or their own custom algorithms on Unity's environments.Does the toolkit send any data back to Unity?Yes - the Unity package includes in-editor analytics to improve developer experience, governed by Unity's general privacy policy.
Yes - the Unity package includes in-editor analytics to improve developer experience, governed by Unity's general privacy policy.Which RL and imitation learning algorithms come built in?PPO, SAC and MA-POCA for reinforcement learning (including self-play and cooperative/competitive multi-agent setups), plus BC and GAIL for imitation learning from demonstrations.
PPO, SAC and MA-POCA for reinforcement learning (including self-play and cooperative/competitive multi-agent setups), plus BC and GAIL for imitation learning from demonstrations.Can I use the Unity environments outside the native ML-Agents workflow?Yes - you can wrap a Unity learning environment as a Gym environment or as a PettingZoo environment to plug into other RL libraries.
Yes - you can wrap a Unity learning environment as a Gym environment or as a PettingZoo environment to plug into other RL libraries.Are there example environments to learn from?Yes - the toolkit ships 17+ example Unity environments covering single-agent, cooperative and competitive multi-agent training scenarios.
Yes - the toolkit ships 17+ example Unity environments covering single-agent, cooperative and competitive multi-agent training scenarios.What are similar projects?AirSim, lerobot, carla, Gymnasium, dqn-adaptive-cpu-scheduler, tianshou, DanKS, dopamine
AirSim, lerobot, carla, Gymnasium, dqn-adaptive-cpu-scheduler, tianshou, DanKS, dopamineWhich topics is it in?imitation-learning, unity3d, deep-reinforcement-learning, unity, neural-networks, game-engine, simulation, game-development, reinforcement-learning, c#, gaming, sdk
imitation-learning, unity3d, deep-reinforcement-learning, unity, neural-networks, game-engine, simulation, game-development, reinforcement-learning, c#, gaming, sdk