TempleFlow

Project Logo
TempleFlow Poster

Description

For the final project of csci1470 (Deep Learning), our team chose option two from the assignment handout and attempted to solve a new problem using deep learning. More specifically, we decided to create a deep learning program which learns how to play a “never-before-attempted” game.

Our game of choice was the popular mobile game Temple Run 2 originally created by Imangi Studios. Our AI player uses a transfer learned convolutional neural network (CNN) to classify real time screenshots of the game into six categories: turn_left, turn_right, jump, slide, lean_left, and lean_right. Our player then takes the predicted action and uses a python library, called pyautogui, to run the corresponding command in our Android game emulator (BlueStacks).

We trained our model by collecting and labeling around 13,000 images (over multiple training trials) and were able to get a test accuracy greater than 98%. Additionally, our model can correctly predict obstacles at runtime with some limitations (see results and challenges section).

At this final deadline, we can declare success on our target goal (see reflection). While we did not reach our stretch goal, of the AI player being able to navigate the game indefinitely, we were able to satisfy two of our other goals: our base goal and target goal.