MechE PhD Final Oral Defense: Xinhuan (Leo) Sang

TITLE: MACHINE LEARNING FOR REAL-WORLD APPLICATIONS: DYNAMICS MODELING, COLLISION AVOIDANCE, AND HUMAN ACTIVITY RECOGNITION

ABSTRACT: This dissertation investigates the application of machine learning to several real-world engineering problems. Machine learning is effective at extracting structure from large datasets, yet its deployment in practice is constrained by limitations and requirements, including computational budgets and safety. We examine how to address these constraints through three projects: surrogate modeling of quadrotor dynamics and acoustics, model-free collision avoidance via reinforcement learning, and IMU-based human gait and activity recognition. In the first project, our goal is to develop a method that provides real-time, medium-fidelity predictions of quadrotor dynamics and acoustics as a function of the vehicle motion state, using a dataset generated from medium-fidelity CFD-based simulations (CHARM for aerodynamic/dynamic outputs and PSU-WOPWOP for acoustic outputs). The method must run at a high update rate (30 Hz) on commercially available general-purpose hardware. We adopt Gaussian processes because they can incorporate gradient information to improve prediction efficiency and accuracy. In our setting, CFD simulations can be post-processed numerically to convert multiple sampled points into a single training point augmented with multi-dimensional gradient information. Leveraging this property, we substantially reduce the computational cost of Gaussian-process prediction. We further partition the state space into multiple subspaces, train a local model for each subspace, and use a Schurcomplement-based technique to re-inject information from data outside the subspace into the corresponding local model, thereby improving accuracy. Together, these techniques yield a partitioned, gradient-enhancedGaussian-process surrogate that maintains real-time performance while achieving higher predictive accuracy than a standard Gaussian process. In the second project, we develop a modular, model-free collision-avoidance method for continuous-time, continuous-space input-affine systems. Classical control approaches, such as CLF/CBF methods and potential/gradient fields, require an accurate dynamics model. By contrast, deep-learning and reinforcement learning methods typically lack formal safety guarantees and often cannot anticipate all relevant situations during training. We derive a value-function-based HJB equality that establishes a quadratic relationship between the advantage function and the control input, providing a theoretical basis for learning. Building on this result, we design an actor–critic learning algorithm that learns an optimal policy and a value function for each environment element. At runtime, given the set of goals and obstacles, we compose the learned policies and value functions by solving a simple QCQP, thereby enabling real-time obstacle avoidance and goal-directed behavior. In the third project, we design a functional electrical stimulation (FES) system that automatically switches stimulation modes based on the activities and gait of post-stroke patients, aiming to support rehabilitation training. This setting requires an algorithm that prioritizes safety when recognizing patient activity and gait. We use real-time lower-limb motion data collected from IMU sensors as input to two GRU-based deep models. During training, we introduce a customized propensity-based evaluation scheme that encourages medically safer predictions under unknown or out-of-distribution conditions. We also leverage a more stable, safety-oriented activity recognition model to further enhance the robustness of gait recognition. We evaluate the system with nine participants to validate the accuracy and stability of activity and gait predictions, and to assess patient adaptation to the system.

COMMITTEE: ADVISOR Professor Roberto Tron, ME/SE; CHAIR Professor Emma Lejeune, ME; Professor Sheryl Grace, ME; Professor Lou Awad, ME; Professor Sean Andersson, ME/SE