Single candidate optimizer based context-aware human motion prediction and action planning for human-robot collaboration using coupled modular transformer network

This is the authors' abstract. We don't add key points for this paper.

Accurate and timely prediction of human motion is crucial for enabling proactive and safe robotic assistance in Human-Robot Collaboration (HRC) scenarios. This paper introduces a new model, coupled with their modular uncertainty-aware decision transformer network with single candidate optimizer (CMUADTN-SCO) to predict human motion with high accuracy and efficiency. Using a large-scale collection of labeled videos of disassembly tasks processes on 30 FPS, 1920 × 1080, the system has had to use the powerful pre-processing methods of spatial resizing and temporal data augmentation to form variable length motion clips. The framework involves an Enhanced VGG19-Graph Attention Network (E-VGG19-GAN) to extract spatial features and make relational decisions which are then fed through the CMUADTN architecture that consists of a Coupled Modular Neural Network (CMNN) to classify motion and an Uncertainty-Aware Decision Transformer (UNREST) to make future predictions. To enable real-time applicability, the Single Candidate Optimizer (SCO) method enhances hyperparameter tuning, optimizing model responsiveness and efficiency. The proposed method surpasses existing state-of-the-art techniques, achieving 98.4% accuracy, 97.9% precision, 97.7% recall, 97.8% F1-score, and 98.9% AUC. These results confirm the framework’s ability to deliver accurate, confident, and safe collaborative behavior recognition in dynamically evolving HRC environments.