Residual Policy Learning for self-driving robot
For a robotic class named Duckietown, we had robots using a camera to follow a road in a model city. The goal was to improve the existing system. We used a Pure Pursuit controller as a basis but refined it using Residual Policy Learning. Instead of modelizing the whole system using Reinforcement Learning, we modeled only a correction that would be applied on top of the PID controller as to make it better. To achieve this, we adapted https://arxiv.org/abs/1509.02971 and used a variety of pretraining. The markdown report can be found here.