Towards Diverse and Natural Stochastic Human Motion Prediction With a Novel Multiobjective Optimization Framework

Stochastic human motion prediction has garnered significant attention in the realm of human-robot interactions. However, existing generative models for this task often struggle with mode collapse. Even the widely used denoising diffusion models face a trade-off between the diversity and accuracy of the generated future human motion sequences. Additionally, these methods overlook the motion diversity difference present within the same pattern of human motions. To alleviate this issue, we propose a novel method called <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Multiobjective Optimization Framework for stochastic Human Motion Prediction</i> (MOFHMP), to enhance the diversity and accuracy of the predicted human motions simultaneously. This proposed MOFHMP reformulates the whole task as a multiobjective optimization problem, and designs a multiobjective optimization network to enhance the diversity and accuracy of predicted human motions. To optimize the designed multiobjective problem, we propose a novel multiobjective optimization algorithm that integrates a Conditional Variational AutoEncoder (CVAE) generative model, which aims to meet the multiple objective requirements in the designed multiobjective function. It is important to note that the Multiobjective Optimization Process (MOP) works in the sampling process of the proposed MOFHMP method, therefore, MOFHMP can enhance the performance of the stochastic human motion prediction without introducing additional training parameters, which means that the proposed method can be easily incorporated to other generative models. In addition, we demonstrate that the proposed MOFHMP also works for controllable human motion prediction, making it a versatile and efficient approach. Through qualitative and quantitative experiments, MOFHMP achieves state-of-the-art results surpassing the baseline methods, validating its superiority in stochastic human motion prediction.