Reinforcement Learning with Q DeepLearning Networks

Github link : github.com/ani...
This 3 minute video shows our RL agent (car) learning to navigate from point to point using Q Learning - a reinforcement learning technique. Q learning here uses a fully connected deep learning network. Deep learning network predicts Q states based on which 3 actions are taken by the agent - move forward, turn left or turn right .
Car here navigates from 3 points. Point A is on top left intersection with coordinates (346, 737). Once it reaches A it moves to B which is on top right (1330, 727). Once that is achieved, it moves back to B. After conquering that too it will then proceed to point C which on bottom left (70, 70). This loop will continue and car will keep hopping from one point to another.
During its travel car is guided by living penalties and rewards. Penalties are punishments given to car for an undesired motion. Rewards are incentives given when it takes a right action. Penalties in this car are for straying out of road, roaming too close to the wall (boundaries of video frame). Rewards given are for travelling through road and travelling towards the target set.