ppo trained carla demo show and method

Our dream is creating a safe driving system working well under all circumstance, for this purpose, a more intelligent agent is needed. We want to create a mind, instead of a larger or deeper network. Now our team using RL algorithm for self-driving, let the agent learning how to drive totally by itself, here is a demo in Carla simulator.

We use two image stack together as input and three-layer convolution network process the image and another mlp net handle measurement. Most important, we never give the car detail command, like go straight 5m then turn right, instead we use a high-level command, like--turn right. Just like when we driving the car, map system tells us turn left, we have to decide when and how to turn, it is the same to our agent.

We use an algorithm called Proximal policy optimization, plus some trick and running with 12 workers, this demo is generated by an agent training about 30 hours.