Toward Context-Aware Exoskeleton Assistance: Integrating Computer Vision Payload Estimation with a Multi-Metric Optimization Space
Back-support exoskeletons mitigate musculoskeletal strain, yet current systems rely on reactive sensing and lack context-aware assistance modulation. This paper presents a population-derived optimization framework and a predictive vision-based adaptive control strategy. First, we construct a multi-metric optimization space combining electromyography reduction, perceived discomfort, and user preference, revealing a non-linear relationship between payload and optimal assistance from experiments with 12 subjects. Second, we develop a computer vision-based adaptive control leveraging a fine-tuned vision transformer (DINOv2) and depth sensing to estimate payloads prior to lifting, eliminating actuation latency. Validation with an additional 12 subjects demonstrates robust payload estimation (82.41% accuracy). The proposed adaptive strategy reduces peak back muscle activation by up to 23% and improves average offloading by 8.15% over static baselines, without increasing discomfort. These results highlight the benefits of predictive perception and user-centric optimization for enhanced human-exoskeleton interaction.