Title: Assessing Deep Learning Efficacy in Predicting Whole-Body 3D Posture During Dynamic Load Reaching

Abstract

This research investigates the utilization of deep neural networks to forecast full-body human posture during dynamic tasks involving load reaching. The study employed two distinct time-series modeling architectures: bidirectional long short-term memory (BLSTM) and transformer networks. Training data comprised 3D full-body plug-in gait dynamic coordinates collected from 20 healthy male participants with normal weight. Each participant executed 204 load-reaching tasks, varying in load positions as well as lifting and handling techniques.

The models utilized specific inputs to predict body coordinates for the final 75% of the task duration. These inputs included the 3D position of the hand-load, body weight and height, lifting styles (stoop, full-squat, and semi-squat), handling methods (one- and two-handed), and 3D coordinate data representing body posture during the initial 25% of the task. To enhance prediction accuracy, the study introduced a novel optimization method that enforces constant body segment lengths via a new cost function.

Results demonstrated that this new cost function reduced prediction errors by roughly 8% for arm models and 21% for leg models. Furthermore, the transformer-based architecture outperformed the BLSTM model in long-term performance, achieving a root-mean-square-error (RMSE) of 41.4 mm and showing approximately 58% greater accuracy. These findings highlight the value of neural networks capable of capturing time-series dependencies in 3D motion frames, offering a distinctive methodology for analyzing and predicting motion dynamics in manual material handling scenarios.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC