This paper presents a study on human motion prediction using recurrent neural networks (RNNs). The authors examine recent work in this area and show that a simple baseline, which does not attempt to model motion, can achieve state-of-the-art performance. They investigate the reasons for this result and analyze recent RNN methods by looking at their architectures, loss functions, and training procedures. The authors propose three changes to standard RNN models for human motion, resulting in a simple and scalable RNN architecture that achieves state-of-the-art performance on human motion prediction. The paper focuses on short-term prediction, which is most relevant for visual tracking. The authors analyze the reasons for poor performance of recent methods on this task by examining factors such as network architectures and training procedures. They propose a sequence-to-sequence architecture with residual connections that models first-order motion derivatives, resulting in smooth and more accurate short-term predictions. The authors also explore the benefits of training a single model to predict motion for multiple actions, as opposed to building action-specific models. They find that training on a large dataset of human motion, such as the Human 3.6M dataset, improves performance. They also find that providing high-level supervision in the form of action labels improves performance, although an unsupervised baseline is still competitive. The authors compare their method to previous work and show that their approach outperforms previous methods in terms of both short-term and long-term motion prediction. They also show that their method produces smooth, continuous predictions in the short term, while still achieving plausible motion in the long term. The authors conclude that their method is simple, scalable, and effective for human motion prediction using RNNs.This paper presents a study on human motion prediction using recurrent neural networks (RNNs). The authors examine recent work in this area and show that a simple baseline, which does not attempt to model motion, can achieve state-of-the-art performance. They investigate the reasons for this result and analyze recent RNN methods by looking at their architectures, loss functions, and training procedures. The authors propose three changes to standard RNN models for human motion, resulting in a simple and scalable RNN architecture that achieves state-of-the-art performance on human motion prediction. The paper focuses on short-term prediction, which is most relevant for visual tracking. The authors analyze the reasons for poor performance of recent methods on this task by examining factors such as network architectures and training procedures. They propose a sequence-to-sequence architecture with residual connections that models first-order motion derivatives, resulting in smooth and more accurate short-term predictions. The authors also explore the benefits of training a single model to predict motion for multiple actions, as opposed to building action-specific models. They find that training on a large dataset of human motion, such as the Human 3.6M dataset, improves performance. They also find that providing high-level supervision in the form of action labels improves performance, although an unsupervised baseline is still competitive. The authors compare their method to previous work and show that their approach outperforms previous methods in terms of both short-term and long-term motion prediction. They also show that their method produces smooth, continuous predictions in the short term, while still achieving plausible motion in the long term. The authors conclude that their method is simple, scalable, and effective for human motion prediction using RNNs.