Godot Native Reinforcement Learning 0.4.0 Misc 4.5 Community

Submitted by user minigraphx; MIT; 2026-07-19

Reinforcement learning for Godot 4.5+ with NATIVE ncnn inference
— statically linked C++, no C#/.NET, no external runtime.

Train with the standard godot-rl (godot_rl_agents) Python stack; deploy native on web/WASM, console, mobile, desktop, and edge — targets an ONNX/.NET pipeline can't reach.

▶ TRY IT IN YOUR BROWSER (no install): https://minigraphx.github.io/godot-native-rl/
The demo launcher as a single-threaded WASM build. Trained agents run on native ncnn with no Python at runtime.
Pick "Evolution Lab" and the browser tab TRAINS a neural net in front of you.

▶ Watch a demo: https://youtu.be/Cud1gvbjg0I

WHAT YOU GET
• A GDExtension that runs trained policies natively via ncnn — one NcnnRunner node, no runtime dependency.
• A godot_rl_agents-compatible training bridge: train with the stock godot-rl Python package, convert to ncnn, deploy.
• Declarative reward authoring (Signal→Reward + RewardBuilder) and a sensors library (raycast, relative-position, camera/pixels, grid, navmesh, entity/attention).
• In-engine ES training (ESTrainer) — evolve a policy with NO Python, NO socket, NO backprop; runs on every target including web.
• Deploy extras a Python-server framework can't match: web/WASM, INT8 quantization, async + batched crowd inference, LOD policy switching, recurrent/LSTM state, continuous + multi-discrete actions.

EXAMPLES (all ship a trained net and run standalone — browse them in the launcher above or the examples/ folder)
Chase the target • Seek & Avoid • Rover 3D • Ball Chase (SAC) • Fly By (PPO continuous) • Hide & Seek (multi-policy self-play) • Coop Collect (MA-POCA) • Quadruped Walk / Hurdles / Jump • Hexapod • 3DBall & GridWorld (Unity-parity) • Visual Chase (CNN from pixels) • Memory Chase (LSTM) • plus the live Evolution Lab.

INSTALL
Enable the plugin, then open the demo launcher (F5).

Full guide:
https://github.com/minigraphx/godot-native-rl/blob/main/docs/guide/getting-started.md

LINKS
• Source & README: https://github.com/minigraphx/godot-native-rl
• Guides (getting started, running examples, training, deploying, sensors): https://github.com/minigraphx/godot-native-rl/tree/main/docs/guide
• ncnn vs ONNX Runtime — honest decision guide: https://github.com/minigraphx/godot-native-rl/blob/main/docs/ncnn_vs_onnx.md
• Issues / questions: https://github.com/minigraphx/godot-native-rl/issues

Pre-1.0: the API and wire protocol may still change between minor versions.
MIT licensed.
You don't need a license to run Models in your games.

Plugin zip link: https://github.com/minigraphx/godot-native-rl/releases/download/v0.4.0/godot-native-rl-addon-v0.4.0.zip
Demos zip link: https://github.com/minigraphx/godot-native-rl/releases/download/v0.4.0/godot-native-rl-examples-v0.4.0.zip