AI learns CTR: Crash Cove PB in 1'52"66 (full course)

This is the result of a Convolutional Neural Network using reinforcement learning over thousands of attempts and hundreds of hours to drive on Crash Cove. The AI reached its objective, which was to drive around the track without wasting too much time, in a consistent manner. It is already fast enough to beat the CPUs on easy mode (not every time). I suggest watching it at 2x speed. I am working on improving the AI and expanding its capabilities. -- Watch live at   / justin_zimmermann   ----------- About PentAI ------------- PentAI is taking the work of @TheRedhotbr on Crash Bash and adapting it to CTR, Crash Cove. Also using Deep Q Learning, with a Convolutional Neural Network. Reward function is proportional to the speed and whether the lap is progressing. -50 points for falling into the water, -0.1 per frame for touching a wall. The AI has as input the 20*20 pixel image of what's in front of it, as a grayscale image (black if undrivable/bad, white if drivable/good). The inputs shown in red are random, while the inputs in light blue are those determined by the Neural Network. Each attempt ends when Penta goes backwards too much, or when it finishes the course. The average reward is over the last 100 attempts.