LongGen-HMP: A Generative Real-Time Framework for Human Motion Prediction via Residual Refinement

Abstract Accurate prediction of human motion using pose information for predicting human trajectories is essential in applications in collaborative environments involving human–robot or human-machine interaction, enabling proactive collision avoidance and a safer workplace. However, forecasting realistic human motion remains challenging due to the high dimensionality, nonlinear dynamics, and the noisy nature of human motion. In this work, we propose a encoder-decoder based Conditional Generative Adversarial Network (cGAN) framework for human motion prediction that generated future pose trajectories from their respective historical observations. The framework employs Feature-wise Linear Modulation (FiLM) to condition both the generator and discriminator on action semantics, enabling action-specific motion generation. The generator learns a distribution over plausible motion trajectories while ensuring spatial and temporal realism. To improve the realism and physical plausibility of the generated trajectories, the training objective incorporates the skeletal constraints together with frequency-aware regularization, preserving both kinematic consistency and the natural temporal rhythms of human motion. An auxiliary Residual Refinement Network (RRN) further reinforces the output by capturing complex inter-limb dependencies. Experimental evaluation on the NTU RGB+D dataset demonstrates that the proposed framework generates smooth, physically consistent, and temporally coherent motion trajectories while maintaining real-time inference capability, making it suitable for safety-critical human–robot collaboration and digital twin applications.