Predicting Human Postures for Manual Material Handling Tasks Using a Conditional Diffusion Model

Predicting workers' body postures is crucial for effective ergonomic interventions to reduce musculoskeletal disorders (MSDs). In this study, we employ a novel generative approach to predict human postures during manual material handling tasks. Specifically, we implement two distinct network architectures, U-Net and multilayer perceptron (MLP), to build the diffusion model. The model training and testing utilizes a dataset featuring 35 full-body anatomical landmarks collected from 25 participants engaged in a variety of lifting tasks. In addition, we compare our models with two conventional generative networks (conditional generative adversarial network and conditional variational autoencoder) for comprehensive analysis. Our results show that the U-Net model performs well in predicting posture similarity [root-mean-square error (RMSE) of key-point coordinates = 5.86 cm; and RMSE of joint angle coordinates = 13.67 $^{\circ }$ ], while the MLP model leads to higher posture variability (e.g., standard deviation of joint angles = 4.49 $^{\circ }$ /4.18 $^{\circ }$ for upper arm flexion/extension joints). Moreover, both generative models demonstrate reasonable prediction validity (RMSE of segment lengths are within 4.83 cm). Overall, our proposed diffusion models demonstrate good similarity and validity in predicting lifting postures, while also providing insights into the inherent variability of constrained lifting postures. This novel use of diffusion models shows potential for tailored posture prediction in common occupational environments, representing an advancement in motion synthesis and contributing to workplace design and MSD risk mitigation.