A leaner, open-source alternative to Unity ML Agents

What it does Godot RL Agents connects Godot Engine games to Python reinforcement-learning libraries through a set of framework wrappers. You can train NPCs or automated play-testers using StableBaselines3, Sample Factory, Ray RLLib, or CleanRL, then ship the results back into a Godot runtime. The project is fully MIT-licensed and includes built-in sensors for 2D and 3D observation spaces.

The interesting bit The authors acknowledge they are following in the footsteps of a well-known proprietary engine toolkit, but their stated goal is to shed the bloat and keep the codebase compact, concise, and hackable. They also support in-editor training inside Godot 4 and offer experimental ONNX export so a trained agent can run inside the engine without dragging a Python interpreter along.

Key highlights

• Wrappers for four major RL frameworks: StableBaselines3, Sample Factory, Ray RLLib, and CleanRL
• Memory-based agents via LSTM or attention interfaces
• Built-in AI sensors for augmenting agent perception in 2D and 3D environments
• In-editor training inside the Godot 4 .NET editor without exporting a game executable
• Experimental ONNX export for deploying trained models natively in Godot

Caveats

• ONNX export is experimental and requires aligning your Godot project’s .csproj and .sln files with the separate plugin source
• Using the default StableBaselines3 script, the Godot environment will freeze periodically while the model updates; the authors note this is expected
• Pre-built test examples are only available for Linux and Windows

Verdict Indie developers and RL researchers looking for a royalty-free, open-source alternative to proprietary engine toolkits will find a pragmatic bridge here. Teams wanting a polished, first-party toolchain with dedicated IDE integration should expect a more hands-on, community-supported experience.