Machine learning-enabled wearable sensing and active exoskeleton control for enhancing human mobility and safety in dynamic environments

Human mobility and safety are increasingly challenged by aging demographics, demanding occupational environments, and unpredictable terrain conditions. In recent years, wearable sensing technologies and active exoskeletons have emerged as transformative tools for enhancing human mobility, reducing musculoskeletal disorders, and mitigating injury risks across diverse real-world settings. From industrial workplaces such as construction sites that demand prolonged standing, squatting, and kneeling to unstructured everyday environments characterized by unpredictable terrain and rapid postural transitions, it has been demonstrated that the integration of real-time sensing with exoskeleton control can improve biomechanical efficiency, reduce injury rates, and facilitate movement assistance. However, several critical research challenges remain unaddressed, including robust real-time activity detection, effective exoskeleton controllers for different gait tasks and environmental conditions, and unified evaluation frameworks for safety, ergonomics, and user experience across varied use cases. To bridge these gaps, this dissertation integrates multi-modal wearable sensing, adaptive exoskeleton control strategies, and extensive experimental validation, contributing to advancements in human-centered wearable assistive robotics. A core component of this dissertation focuses on real-time gait recognition and posture tracking in both industrial and daily-living contexts. Through an optimized inertial measurement unit (IMU)-based sensing approach, we demonstrate that even a minimal set of body-worn sensors enables highly accurate classification of fundamental movement patterns, including walking, squatting, kneeling, and transitional activities. Using machine learning algorithms applied to low-latency sensor streams, the classification design achieves accuracies exceeding 95\% across varied conditions, and thus highlights the feasibility of compact, efficient sensing systems for practical deployment. The sensing framework is further extended to detect and respond to unexpected slips with wearable IMUs, where real-time exoskeleton torque adjustments at the bilateral knee joints significantly reduce fall risk and recovery time. Additionally, a hybrid neurophysiological sensing paradigm is introduced, integrating electroencephalogram (EEG)-based neural signals with IMU-based kinematics to predict sit-to-stand transitions before the physical execution of movement. This fusion of neural intent detection and inertial sensing leads to a measurable reduction in detection latency, enabling smoother and more responsive exoskeleton assistance that minimizes user exertion during daily activities. This dissertation further extends to exoskeleton control for unstructured and deformable terrains, focusing on human gait over soft, uneven surfaces such as sand. Walking on yielding substrates imposes heightened metabolic demands due to increased energy dissipation and unstable ground reaction forces. To address this challenge, a model predictive stiffness-based exoskeleton controller is introduced to dynamically adjust knee joint torque in response to terrain variability and user intent. Through an integrated analysis of kinematics, kinetics, and muscle activation, the exoskeleton control approach is shown to effectively reduce lower-limb strain and overall effort, allowing users to traverse deformable terrain with enhanced efficiency and reduced biomechanical load. Beyond fundamental locomotion assistance, this dissertation further investigates the role of exoskeleton control in posture stability and skill-based tasks, particularly in environments with increased physical and cognitive demands. To examine these effects, experimental setups incorporating virtual- and mixed-reality (VR/MR) environments are designed to simulate elevated construction settings with task constraints such as welding and precision-based manual work. By analyzing center-of-pressure stability, muscle activation patterns, and joint kinetics, the influence of exoskeleton torque modulation on stance control and kneeling balance is quantified, particularly at high elevations where psychological factors such as fear of falling can negatively impact performance. Multi-subject experiments reveal that exoskeleton support significantly improves postural stability and reduces lower-limb fatigue, demonstrating its potential for physically intensive professions where sustained joint loading and instability are prevalent. To further validate the long-term sustainability of the knee exoskeleton’s performance benefits, we conduct an extended evaluation study involving real-life construction tasks. Additionally, an open-source multimodal dataset is released to facilitate future research and benchmarking in the field. Collectively, the findings of this dissertation contribute to a holistic framework for next-generation wearable robotic systems, beginning with high-fidelity real-time sensing of human motion, progressing to adaptive control strategies that enhance stability and mobility, and culminating in human-centered validation through extensive experimental studies. Across multiple human-subject trials, it is established that minimal sensor configurations can reliably detect gait phase and activity transitions, that rapid slip detection and torque-assisted recovery mitigate fall risk, and that advanced model predictive controllers are crucial for enabling efficient movement over irregular and unpredictable surfaces. By seamlessly integrating wearable sensing, intelligent control, and biomechanical insights, this dissertation lays the groundwork for scalable, real-world deployment of exoskeleton technologies that enhance mobility, reduce injury risks, and improve human performance in both occupational and daily activity environments.