Deep QWOP Learning

Abstract

We apply a deep learning model to the QWOP flash game, which requires control of a ragdoll athlete using only the keys “Q”, “W”, “O”, and “P”. The model is a convolutional neural network trained with Q-learning. By training the model with only raw pixel input, we show that our model is capable of successfully learning a control policy associated with playing QWOP. This model was successfully applied to a non-deterministic control environment in the form of a ragdoll physics flash game.