Im a student that is trying to finish my degree. For that im trying to learn about Reinforcement Learning. Due that Im implementing DQN on a Gazebo-Ros environment where I use a Turtlebot3 Burger model and a Square as a Goal. I wanted to apply CNN nets to this algorithm (I tried with Lidar sensors and the algorithm seems to work). The problem with the CNN nets its that they are not learning. I dont know whats the reason. The images that im passing to the net are images (64x64x3) from top view of the map something like this: Top view of map.