Autonomous Driving based on Reinforcement Learning in Unity

This is the Final Year Project for my Bachelors in Nanyang Technological University. The car model used is a Donkey Car, an open source DIY self driving platform for small scale RC cars. However, only imitation learning is officially supported on their platform. This project makes use of Unity's ML-Agents package for reinforcement learning, along with MMSegmentation for semantic segmentation.

Features

• Domain Randomization Lighting, wall colours, and terrain textures are constantly changing during simulation
• Semantic Segmentation Option to train a policy with the output of a semantic segmentation model. A SegFormer-B0 model was trained with synthetic data generated from Unity. Note: Testing in Unity can only be done with ground truth masks as many models, including SegFormer, are not supported by Unity's Barracuda package.
• Obstacle Avoidance Scene with obstacles in the road for training obstacle avoidance.

Setup

Unity 2020.3.11f1

$ git clone https://github.com/MrOCW/Autonomous-Driving-RL-Unity
$ cd Autonomous-Driving-RL-Unity
$ git clone https://github.com/MrOCW/ml-agents ml-agents-2.1-dev

Local installation for development

$ cd ml-agents-2.1-dev
$ pip3 install -e ./ml-agents-envs
$ pip3 install -e ./ml-agents

Install ML-Agents in the Unity Editor by following the installation guide