Toward Context-Aware Exoskeleton Assistance: Integrating Computer Vision Payload Estimation with a User-Centric Optimization Space

TL;DR

A user-centric optimization framework and a vision-based adaptive control strategy for industrial BSEs are introduced, demonstrating that pre-lift environmental perception and user-centric optimization significantly enhance physical assistance and human-robot interaction in industrial settings.

Abstract

Back-support exoskeletons (BSEs) mitigate musculoskeletal strain, yet their efficacy depends on precise, context-aware modulation. This paper introduces a user-centric optimization framework and a vision-based adaptive control strategy for industrial BSEs. First, we constructed a multi-metric optimization space, integrating electromyography reduction, perceived discomfort, and user preference, through baseline experiments with 12 subjects. This revealed a non-linear relationship between optimal assistance and payload. Second, we developed a predictive computer vision pipeline using a Vision Transformer (DINOv2) to estimate payloads before lifting, effectively overcoming actuation latency. Validation with 12 subjects confirmed the system's robustness, achieving over 82% estimation accuracy. Crucially, the adaptive controller reduced peak back muscle activation by up to 23% compared to static baselines while optimizing user comfort. These results validate the proposed framework, demonstrating that pre-lift environmental perception and user-centric optimization significantly enhance physical assistance and human-robot interaction in industrial settings.

Figures (7)

• Figure 1: Overview of the CVAC pipeline (a): images captured by the RPI are transmitted to the cloud workstation for object detection and payload estimation, and the resulting predictions are returned to the RPI for serial communication with the exoskeleton. Panels (b) and (c) illustrate the architectures of the payload-estimation models; MLP denotes Multilayer Perceptron.
• Figure 2: The back support exoskeleton Ding (a) and validation experiments setting (indoor environment) (b).
• Figure 3: Block diagram of the low-level control architecture. The mid-level assistance law combines gravity support and velocity shaping to generate $\tau_{\mathrm{cmd}}$. The CVAC-driven gain $k_{pyl}^{j}$ scales the desired torque, and the SEA torque controller (PID + DoB) compares the spring-deflection torque estimate with $\tau_{\mathrm{adj}}$ to command motor current and regulate the measured hip torque.
• Figure 4: EMG, Discomfort, Preference, and Total Representation Functions from both Baseline (I) and Validation (II) tests. EMG-ORF (Optimization Representation Functions) and EMG-VRF (Validation Representation Functions) are divided into mean/peak values for back muscles, leg muscles, and total muscle contributions. Similarly, T-ORF and T-VRF are evaluated for each back/all mean/peak combination. For EMG functions, the scale on the right indicates the percentage reduction relative to the no-exoskeleton condition, while for other functions, it represents the normalized value. LW, MW, and HW denote low, medium, and heavy weight conditions. Assistance strategies include LA, MA, and SA for light, medium, and strong assistance in Baseline tests, while StA and AdA refer to static and adaptive assistance in Validation tests.
• Figure 5: Gradient-based optimization of $J_{tot}$ and fitting process of the optimization points. Arrows show the direction to follow for performance improvement, with red dots indicating the maximum O-RF values.