Unity ML-Agents — Deep RL for Game NPCs

Unity ML-Agents — Deep RL for Game NPCs

Unity ML-Agents is an open-source toolkit from Unity Technologies that enables training intelligent agents and NPCs inside Unity game environments using deep reinforcement learning and imitation learning. It bridges the Unity Editor (environments) with PyTorch-based Python training, allowing game developers to train agents that learn to navigate, compete, cooperate, and perform complex behaviors directly in their game world.

Key Features

• Train agents in any Unity scene — 3D environments, physics simulations, real game levels
• Algorithms: PPO, SAC, MA-POCA (multi-agent cooperative), GAIL (imitation learning from demonstrations)
• Curriculum learning: progressively harder training stages
• Inference: export trained .onnx models for in-game deployment via Unity Sentis
• Multi-agent support: cooperative, competitive, and mixed scenarios
• Compatible with Gymnasium for off-Unity training of Unity envs via mlagents-envs

Quick Start

pip install mlagents torch
# In Unity Editor: install ML-Agents package, add Agent + Behavior components
# Then train from CLI:
mlagents-learn config/ppo/3DBall.yaml --run-id=3DBall_run1

# Use a trained Unity env from Python
from mlagents_envs.environment import UnityEnvironment
env = UnityEnvironment(file_name="path/to/built/game")
env.reset()
behavior_name = list(env.behavior_specs)[0]
decision_steps, terminal_steps = env.get_steps(behavior_name)

npx ai-supply add unity-ml-agents-game-ai

Curated mirror of the open-source Unity ML-Agents (Apache-2.0). Get it from the source.