EMG-Based Human-in-the-Loop Bayesian Optimization to Assist Hip-Centric Activities

This dissertation presents a practical method for personalizing hip exoskeleton assistance using surface EMG-based human-in-the-loop optimization, cutting tuning time from hours to minutes while preserving assistance quality. We show that processed EMG provides a reliable objective for rapid personalization, enabling convergence within typical clinical sessions. The research progresses from simulation studies revealing fundamental controller-hardware gaps to experimental validation across three activities. In leg swinging (n=8), EMG-based optimization reduces muscle activity by 15-17\% with <15 seconds of steady data per trial. In squatting (n=4), the method completes tuning in 4 minutes 40 seconds and yields 21\% lower metabolic cost with 17\% lower EMG. In walking (n=11), a multi-objective formulation balancing EMG and user preference identifies personalized controllers in 11-12 minutes, reducing metabolic cost by 14.9\% while improving perceived exertion by 25-45\%. Three technical innovations enable this speed: (i) a signal-enhancement pipeline combining Hankel decomposition, Bayesian regularization, and optimized smoothing that improves composite EMG quality metrics by 108\%; (ii) machine-learning-guided initialization from anthropometric measurements that reduces convergence time by 26.5\% and improves final performance by 9.98\%; and (iii) heteroscedastic Gaussian process surrogates with Expected Hypervolume Improvement that capture input-dependent noise, improving predictive accuracy by 23-31%. Supporting investigations establish practical design principles. Systematic evaluation across 12 participants performing 30 conditions each reveals that simple amplitude summation provides the most reliable EMG-metabolic correlation (r=0.762), challenging assumptions about complex feature necessity. Simulation studies comparing 12 controller architectures demonstrate that phase-adaptive impedance achieves 55.5\% mechanical power reduction with minimal parameters, while Bezier profiles reach 62.9\% reduction at higher implementation cost. These results establish EMG-based HIL as a clinically feasible approach to exoskeleton personalization, validated across 23 healthy adults. The framework employs compact controller parameterizations (4-8 parameters) suitable for real-time optimization, with transparent cost functions and traceable convergence. Limitations include healthy-adult validation, electrode placement sensitivity, and restricted activity scope. Extensions should address clinical populations, online adaptation to fatigue, and broader task coverage.

Author dc:creator

• Salvador Echeveste (23292001)

• Engineering
• Robotics

Statement dc:rights

• In Copyright
• Open Access after 2028-01-01

DOI dc:identifier

https://doi.org/10.25417/uic.31451620.v1

OAI identifier oai:identifier

oai:figshare.com:article/31451620