This paper introduces two deep neural network architectures for action-conditional video prediction in Atari games. The models are designed to predict future frames based on previous frames and control actions. The architectures consist of encoding, action-conditional transformation, and decoding layers using convolutional and recurrent neural networks. The models are evaluated on the Arcade Learning Environment (ALE) and show the ability to generate visually realistic frames for up to 100-step predictions. The paper is the first to make and evaluate long-term predictions on high-dimensional video data conditioned by control inputs. The first architecture uses feedforward encoding, which processes a fixed history of previous frames through convolutional layers to extract spatio-temporal features. The second architecture uses recurrent encoding, which processes one frame at a time and uses an LSTM to model temporal dynamics. Both architectures include an action-conditional transformation layer that uses multiplicative interactions between encoded features and control variables to generate predictions. The decoding layer maps the predicted high-level features into pixel values. The models are trained using a curriculum learning approach, starting with short-term predictions and gradually increasing the number of steps. The models are evaluated on the quality of predicted frames and their usefulness for control. The results show that the models outperform baselines in both qualitative and quantitative measures. The models are also able to distinguish between controlled and uncontrolled objects, indicating that they learn disentangled representations of the environment. The paper also evaluates the usefulness of the predicted frames for control in two ways: by replacing the emulator's frames with predicted frames in a DQN controller and by using the predictions to improve exploration in DQN. The results show that the models significantly improve the performance of DQN in several games. The models are also able to handle complex interactions between objects and predict long-term dependencies in the environment. The paper concludes that the proposed architectures are effective for action-conditional video prediction in Atari games and have the potential to generalize to other vision-based reinforcement learning problems.This paper introduces two deep neural network architectures for action-conditional video prediction in Atari games. The models are designed to predict future frames based on previous frames and control actions. The architectures consist of encoding, action-conditional transformation, and decoding layers using convolutional and recurrent neural networks. The models are evaluated on the Arcade Learning Environment (ALE) and show the ability to generate visually realistic frames for up to 100-step predictions. The paper is the first to make and evaluate long-term predictions on high-dimensional video data conditioned by control inputs. The first architecture uses feedforward encoding, which processes a fixed history of previous frames through convolutional layers to extract spatio-temporal features. The second architecture uses recurrent encoding, which processes one frame at a time and uses an LSTM to model temporal dynamics. Both architectures include an action-conditional transformation layer that uses multiplicative interactions between encoded features and control variables to generate predictions. The decoding layer maps the predicted high-level features into pixel values. The models are trained using a curriculum learning approach, starting with short-term predictions and gradually increasing the number of steps. The models are evaluated on the quality of predicted frames and their usefulness for control. The results show that the models outperform baselines in both qualitative and quantitative measures. The models are also able to distinguish between controlled and uncontrolled objects, indicating that they learn disentangled representations of the environment. The paper also evaluates the usefulness of the predicted frames for control in two ways: by replacing the emulator's frames with predicted frames in a DQN controller and by using the predictions to improve exploration in DQN. The results show that the models significantly improve the performance of DQN in several games. The models are also able to handle complex interactions between objects and predict long-term dependencies in the environment. The paper concludes that the proposed architectures are effective for action-conditional video prediction in Atari games and have the potential to generalize to other vision-based reinforcement learning problems.