Live demo of CADDY trained on Atari Breakout. Starting from an initial frame, our model receives an action input at each step and predicts the successive frame. Use the buttons at the bottom of the figure to specify the current action input, the 'reset' button to restart from the beggining and the 'new' button to start from a different frame.
To guarantee better compatibility across devices, this demo runs on CPU, while our complete model runs on GPU in real time. Moreover, running in browser required minor modifications to the model which may cause losses in performance. The versions of the model trained on the BAIR and on the Tennis dataset could not be included due to the limitations on the size of the submission files.