UP TO DEATH is an adventure hypercasual game created by our friend Rachit Vasudeva [itch.io]
From the game's page [link]:
"This, is UP TO DEATH, an adventure hypercasual game in which you have to rescue CUBEY, from being collided with the horde of spikes at the top. Collect as much stars as you can, create a monster score and compete with your friends."
Inspired by a video from ClarityCoders on using an Image Classifier model to create a bot to play Fall Guys, we decided to give that technique a try and apply it on UP TO DEATH.
The base of the source code is a fork of ClarityCoder's repo. We have made many changes in the scripts to suit our needs in detecting the game screen anywhere on the display, creating scripts to create training data while the human (us) is playing and another script to process bad runs, associate keystrokes with screenshots
The technique is mainly to associate a keystroke to a screen capture in the game and hope the model picks up the pattern, which is visible in the demo run trained on a dataset of 97 thousand images. We were able to a recorded high score of 14 after many tries
We were also observed that the model exhibited our gameplay styles and sometimes that showed up uniquely such as Ya-s-h's "wiggling left and right" tactic
demo-run-14-highscore.mp4
you can find the colab file used to train the model on their video: FastAI learns to play Fall Guys