LearnToDrive

Machine learning game using Unity ML agents (tensorflow). The aim is to make the car learn to drive without hitting walls

It's a sort of self driving car simulator, only using "sensors" for driving, real car use also computer vision with cameras (see improvement ideas)

Usage

In order to make this project work, i recommend following tutorials to setup Unity ML-Agent https://github.com/Unity-Technologies/ml-agents

Inputs variable

x value of relative position between point to reach and agent
z value of relative position between point to reach and agent

Check if hitting something at angles (2 origins) : 0, 45, 90, 135, 180, 110, 70

Output variable

value to move on x axis
rotation

Game

Videos

Preview of 15 min trained agent - v0.1

Version 0.1

Preview of trained agent - v0.5 – 100 000 steps

Version 0.5

Preview of trained agent - v0.6 ~ 100 000 steps

Version 0.6

Preview of trained agent - v0.9 - 678 000 steps - Handling speed now (not very well)

Version 0.9

Improvement idea

• Use better assets (real car, real world : roads...)
• Use computer vision + actual kind of sensors
• Build small car with Raspberry PI and sensors that use the model