Mind Mesh

Mind Mesh is an experimental AI racing project where neural networks learn to drive formula-style cars on procedurally generated tracks. Built entirely in Unity with C#, it features realistic car physics and evolutionary neural networks trained through natural selection.

🚗 How It Works

Each generation spawns 100 AI agents. They receive input from:

• 15 raycasts spread over 180° (for spatial awareness)
• Vehicle speed
• Wheel slip (both forward and sideways for each wheel)
• Track direction

Outputs are:

• Steering direction
• Throttle/Brake (via a single node: top half = throttle, bottom half = brake)

Agents evolve through generations using mutation and selection. The best-performing driver (distance + low impact or no crash) passes its traits to the next generation.

🧠 Neural Network

The architecture is a multi-layer feedforward network built from scratch in C#. It supports:

• Randomized weights and biases
• Mutation-based evolution
• Activation via sigmoid function
• Output decisions per frame
• Future: competitive training with collisions between agents

🛞 Physics & Control

Cars use Unity’s Rigidbody system and Wheel Colliders, simulating:

• Traction & friction
• Suspension
• Engine torque and directional steering
• Speed-based throttle/brake handling

A brake override system was added after early agents pressed throttle + brake simultaneously. This significantly improved lap time efficiency.

🧪 Results

• After ~50 generations, agents completed laps consistently without crashes.
• They learned subtle driving behaviors like slowing before corners and staying right on the track due to spawn conditions.
• New additions include “reverse” and “idling” death conditions to prevent cheating/stalling.

I love how realistic the agents became. They started driving like maniacs, then learned restraint—just like we do.

🧰 Tools Used

• Unity 2022.3 LTS
• C#
• Blender (car modeling + procedural track generation using Geometry Nodes)

🛣️ Next Steps

• Enable full collision between agents → true competitive AI racing
• Export a public web demo