How sookeyy-12/neuralnitro-ai works
Project type and purpose
• This is a Unity-based full-stack simulation & training environment (“Neural Nitro”) for reinforcement learning (RL) agents in a racing‐car scenario.
• The Unity project (C# scripts + Assets + Scenes) provides the simulation, and Unity’s ML-Agents toolkit connects to an external Python trainer (via gRPC) to drive training/evaluation loops.
File structure
• Top-level folders:
• Assets/… → Unity assets: Scenes, Prefabs, Models, Materials, Scripts, Textures, ML-Agents config, ONNX model files under Assets/Models.
• Config/ & Curriculum/ → YAML configs and pretrained checkpoints for Python trainer.
• Packages/ → manifest.json for Unity packagemanager (incl. com.unity.ml-agents).
• ProjectSettings/ → Unity project settings.
• results/ → per‐experiment training outputs: ONNX models, .pt checkpoints, run_logs (timers, training_status).
• Key patterns:
• MVC-like separation: Unity Scenes (View), C# Scripts (Controller), ScriptableObjects/configs (Model).
• External ML service (Python) drives learning, writes models back into Assets/Models for inference in Unity.
• Build/deploy: Uses Unity Editor + ML-Agents package. Training is orchestrated by executing Unity in “training mode” connecting to Python.
README for additional insights
• Describes observation (RayPerceptionSensors, speed), action spaces (Throttle, Steer, Brake), reward structure.
• Instructs how to fix ML-Agents package paths in manifest.json.
• Highlights use of PPO (and other algorithms) via ML-Agents.