Publications

The Learning from Demonstration (LfD) literature is growing at a pace that makes comprehensive coverage increasingly difficult. This Science and Technology Watch column presents a structured one-year study of the LfD field, curating and synthesizing key contributions across imitation learning, policy learning, and robot programming by demonstration. The work identifies emerging trends, highlights representative results, and outlines open challenges — serving as a navigational aid for researchers and practitioners entering or tracking this rapidly evolving area.

Citation: A. Billard, R. Detry, N. Figueroa, M. Foriest, D. Lee and K. Yao, "Surviving the Paper Deluge: A One-Year Study in Learning From Demonstration [Science and Technology Watch]," in IEEE Robotics & Automation Magazine, vol. 33, no. 2, pp. 205-211, June 2026, doi: 10.1109/MRA.2026.3682522.

Proposes a deep learning-assisted NDE technique for locating and sizing coating delaminations using ultrasonic guided waves. A transducer sends guided waves into a coated plate; time-domain signals from multiple downstream receivers form a time-space image that feeds into a trained ML model which directly outputs delamination location and size. Numerical simulations show high throughput and accuracy after a one-time training phase, making the approach practical for real-world NDE and SHM scenarios.

Citation: Junzhen Wang, Maximilian Schmitz, Laurence J. Jacobs, Jianmin Qu, Deep learning-assisted locating and sizing of a coating delamination using ultrasonic guided waves, Ultrasonics, Volume 141, 2024, 107351, ISSN 0041-624X, https://doi.org/10.1016/j.ultras.2024.107351.

Proposes a deep learning framework to predict interfacial conditions in coated plates using guided wave measurements. Spring compliance parameters model bond quality; an LSTM network predicts tangential and normal compliance from dispersion curves derived via FEM simulation. For full delamination cases, a CNN processes time-space images from downstream receivers to predict delamination location and size. Both methods demonstrate strong potential for practical NDE and structural health monitoring applications.

Citation: Junzhen Wang, Maximilian Schmitz, Laurence J. Jacobs, and Jianmin Qu, Deep learning-based prediction of interfacial conditions in coated plates using guided waves, Proc. SPIE 12951, Health Monitoring of Structural and Biological Systems XVIII, 129511F (9 May 2024); https://doi.org/10.1117/12.3010200.

Applies machine and deep learning to nondestructively characterize coating thickness and uniformity in layered systems. Finite element simulations generate Lamb wave time-domain signals that are 2D-Fourier-transformed into dispersion maps. For uniform coatings, extracted dispersion features feed ML classifiers for thickness categorization. For non-uniform coatings, a convolutional neural network (CNN) trained on simulated dispersion maps achieves robust classification, with recommendations given on network architecture and evaluation.

Citation: Maximilian Schmitz, Jin-Yeon Kim, Laurence J. Jacobs, Machine and deep learning for coating thickness prediction using Lamb waves, Wave Motion, Volume 120, 2023, 103137, ISSN 0165-2125, https://doi.org/10.1016/j.wavemoti.2023.103137.

Master's thesis investigating machine and deep learning for nondestructive characterization of thin coating quality — thickness and uniformity — in layered systems. Dispersion curves obtained from FEA-simulated Lamb waves are processed via 2D-FFT into frequency–wavenumber maps. ML classifiers handle uniform thickness prediction; two CNN architectures extend this to non-uniform coatings. The work evaluates both networks and provides guidance on architecture selection and practical deployment.

Citation: Maximilian Schmitz, Deep Learning in Ultrasonic Wave Inversion for Thin Coatings, Master’s Thesis, Georgia Institute of Technology, Atlanta, GA, 2022.

Manually tuning controller parameters is fundamental but tedious and potentially unsafe when done on a real system. This paper applies SafeOpt — a Bayesian optimization algorithm — to automatically tune control law parameters while guaranteeing that performance never falls below a safety threshold. A Gaussian process models the performance function, and only parameter configurations with high-probability safe performance are evaluated, starting from a low-performance but safe initial controller.

Citation: Schmitz, Maximilian, Gray, Justin, Oh, Jaeyo, Lu, Yuwei, Kanwar, Bharat (2022). "Gaussian Processes for Automatic Con- troller Gains Tuning in Robotics and Control." (unpublished).

Final project for the Engineering Cybernetics graduate competition in advanced control theory. The paper covers the design of nonlinear state-feedback control algorithms and the numerical optimization of a race trajectory for a single-track vehicle model, balancing control performance with computational feasibility.

Citation: Schmitz, Maximilian, Ruehle, Josias. (2020). "Project Report: Optimize a Single Track Vehicle Model with Non-linear State-Feedback Controller ." (unpublished).