TinyAIDriver

A minimal obstacle avoidance self-driving AI written in C++. The AI is built on neural networks trained with an evolutionary algorithm (aka Genetic Algorithms).

The goal of this project is to have a minimal ANN infrastructure for simple simulations where controllers are involved.

This is far from being optimal in any way. It's meant as a starting point and an educational tool. Practical applications are possible, but for better performance you're better off using LibTorch.

Features

• Neural network-driven self-driving AI (obstacle avoidance)
• Neural Network training and inference in C++ with no external dependencies
• Real-time visualization of the simulation and training progress
• Minimal immediate-mode rendering API based on OpenGL

Building and Running

Prerequisites

• CMake 3.7 or higher
• C++20 compatible compiler
• Git (Git Bash on Windows)

Getting the Dependencies

Required external dependencies are fetched by calling:

./get_externals.sh

Building the Project

./build.sh

The executable will be placed in the _bin directory.

Running the Simulation

./_bin/TinyFreeway

Wait a few seconds while the first batch of networks is trained, then the simulation will start to play, progressively improving over time.

Command-Line Options

• --help: Display help information
• --use_swrenderer: Use software rendering instead of hardware acceleration
• --autoexit_delay <frames>: Automatically exit after a specified number of frames
• --autoexit_savesshot <fname>: Save a screenshot before automatic exit

Controls

The user interaction is limited to tweaking the GUI controls. It's safe to play and see.

Project Structure

• TinyFreeway/: Self-driving AI implementation and simulation
• Common/: Core utilities for SDL2, OpenGL, and ImGui integration
• _externals/: External dependencies

External Dependencies

Graphic display and user interface are implemented using the following libraries:

• SDL2
• ImGui
• GLM
• fmt

Code overview

The actual AI engine, reusable part of this demo, is defined in the following sources:

• TA_EvolutionEngine.h
• TA_SimpleNN.h
• TA_Tensor.h
• TA_TrainingManager.h
• TA_QuickThreadPool.h

SimpleNN and Tensor are the low-level building blocks of the neural network.

EvolutionEngine is responsible for the genetic algorithm that given a population of neural networks and their fitness, produces a new generation of networks.

TrainingManager orchestrates the training process, by calling the evaluation function and passing the results to the EvolutionEngine.

QuickThreadPool is a simple thread pool implementation that allows the training process to be parallelized.

The simulation logic for the synthetic environment is contained in the Simulation class.

In main.cpp, a calcFitnessFn function is defined for TrainingManager, which is responsible for running the simulation with a given neural network and returning its fitness (success score).

The rest of the code is for display and user interface.

Contacts

• You can find me on X with the handle 109mae
• My company: NEWTYPE
• My game studio: OYK Games

License

See the license.txt file for details.

Acknowledgements

This project is a spin-off of Demo9 from the dpasca-sdl2-template repository.