Repetitive history-driven fuzzy Markov networks for long-term human motion prediction

Human pose prediction aims to forecast future motion from historical sequences and plays an important role in autonomous driving, human-machine interaction. Although recent diffusion models have achieved remarkable progress in human motion prediction, diffusion-based human prediction still faces two critical challenges. First, the spatial distribution of multi-level joints is highly dispersed and exhibits substantial variation in value ranges, making it difficult for models to effectively capture spatial heterogeneity across joints. Second, most existing methods apply attention mechanisms that treat all frames and joints uniformly, which limits their ability to model periodic and repetitive motion patterns embedded in historical sequences. This limitation often leads to suboptimal performance, particularly for actions with strong temporal repetition. To address these challenges, we propose the Repetitive history-driven Fuzzy markov Network (RFN). At the core of RHF is the Fuzzy-markov chain noise Generation Module (FGM), which introduces ordered/chaotic and smooth/abrupt descriptors to characterize motion variations. By leveraging neural modules inspired by the conceptual principles of Sugeno and type-1 fuzzy systems-specifically, rule-based semantic partitioning and singleton-like output modulation-our framework translates fuzzy motion attributes into deterministic representations and dynamically calibrates noise distributions across joints. This anatomy-aware mechanism significantly enhances the model's capacity to capture spatially varying motion dynamics. In addition, we design the History Repeat driven Module(HRM), which employs a window-based action modeling mechanism and dual-view temporal modeling to extract both global and local temporal dependencies. A Retrospective Replay for Refinement (3R) mechanism is further incorporated to strengthen the capture of key repetitive motion patterns. Our method outperforms the state-of-the-art models on ADE and FDE metrics, demonstrating the effectiveness of our approach. In particular, our method achieves a significant improvement in MMADE and MMFDE metrics, showcasing the advantages of our model in handling diverse motion prediction tasks.