Toward User- and Task-Adaptive Exosuit Assistance using Perceptually Guided Optimization and Physics-Inspired Machine Learning

Abstract

Wearable robotic devices such as exosuits and exoskeletons are becoming increasingly prevalent across various domains of daily life, including both occupational and clinical settings. As these de- vices are transitionining from the lab into the real world, however, new challenges arise. Crucially among them is the ability to deliver assistance that is appropriate across a variety of users and task spe- cific demands, arising from factors such as differences in anthropometrics, fatigue, motor learning, or individual perception of restriction or support. This thesis is organized around three aims, situated within the broader goal of moving toward assistance that responds dynamically to who is wearing the device, the activity they are doing, and how these demands change over time. First, this work examined whether and how user preference can serve as a viable optimization target for personalizing exosuit assistance during a discrete motor task. To this end,we developed a preference- based Bayesian optimization framework to tune lifting and lowering assistance for a soft back exosuit, and incorporated user-specific just-noticeable differences (JNDs) as perceptual thresholds to adap- tively guide exploration away from indistinguishable comparisons. Applied to lifting, this framework yielded preferred settings that were reliably identified within users and uniquely different between them. It further provided quantitative evidence that participants preferred asymmetric assistance, and showed in simulation that JND-aware sampling improves optimization efficiency under elevated perceptual uncertainty. Building on these findings, we translated this single-session framework into a clinical context, where user needs evolve over the course of recovery. The preference algorithm was adapted into a two-stage, hierarchical protocol, combining a coarse-tuning stage at every visit with a fine-tuning stage at biome- chanical assessments, and embedded within a prospective, registered, single-arm controlled trial of exosuit-augmented physical therapy for low back pain. Tracking preferred settings across treatment visits, preferred assistance was found to evolve asymmetrically over time, with users accepting greater lifting assistance as recovery progressed while preferences for lowering assistance remained relatively stable. The third aim was to develop an approach for estimating joint moments from wearable sensor inputs, providing a task-agnostic control signal and richer biomechanical insights. To this end, this thesis developed a physics-inspired, modular deep neural network for estimating bilateral lower-limb joint moments from sparse kinematic inputs, combining hard structural constraints with soft loss- based regularization. This approach improved generalization across unseen post-stroke study cohorts, in low-data regimes, and on unseen user anthropometrics. A differentiating feature of this design is that, alongside net joint moments, it provides biomechanically meaningful intermediate outcomes, including constituent torque components, ground reaction forces, and centers of pressure, allowing error contributions to be traced back through the prediction pipeline while offering insight into inter- and intra-limb dynamical relationships that pure end-to-end models cannot provide. Together, these contributions advance preference-based personalization and physics-inspired joint moment estimation as complementary foundations for user- and task-adaptive control of soft exosuits.