Autonomous Racing Track Agent
Principle
How to build a simulated racing track driver that would learn to drive based on vision with the intuition of a racing line?
Abstract
VRONA® BOT88 is a self-driving car project that leans toward building an autonomous racing driver model based on plain input vision (no lane and surface detectors preprocessing). It learns to control a Porsche 911 GT3 R on (Barcelona) Catalunya GP racetrack in the notorious high quality racing simulation (sim-racing) title, Project Cars 2™.
The underlying idea is like a racing driver that uses his/her senses (then skills) to perform, the agent should learn in the same context as a driver.
DDPG algorithm has been used with asymmetric actor-critic. This publishing is focused on BOT88’s training phase.
Development made with Python 3.6 – Keras. Computation made on MS Azure virtual machine with Nvidia GPU.
