Name: CARLOS ARTURO NARVAEZ DELGADO
Publication date: 20/03/2026
Examining board:
Name |
Role |
|---|---|
| CAMILO ARTURO RODRIGUEZ DIAZ | Presidente |
| CÍCERO MARTELLI | Examinador Externo |
| MARIANA LYRA SILVEIRA | Examinador Interno |
| MOISES RENATO NUNES RIBEIRO | Coorientador |
Summary: Evaluation of the physical Human–Robot Interaction (pHRI) in upper-limb exoskeletons remains challenging due to anatomical variability, complex motion, and the limitations of conventional sensing when contact forces are spatially distributed and structurally coupled. This master’s dissertation presents the development and experimental assessment of a
sensing module instrumented with Fiber Bragg Grating (FBG) sensors—as a building block toward a future pHRI testbench—based on a 3D-printed, anthropometric forearm-like structure divided into eight contact sections. A reproducible protocol with multiple force levels was implemented, and a data-driven pipeline was adopted to estimate both force location and force magnitude from multichannel optical measurements, combining spatial separability analysis with clustering, supervised learning, point-wise linearization, and spatial interpolation. Spatial separability was verified through unsupervised clustering using statistical feature pairs and triplets, where the best feature pair achieved a cluster purity of 0.933 and the best feature triplets achieved a purity of 1.000, indicating strong correspondence between the discovered clusters and the true contact sections. For supervised estimation, several models were compared, and Random Forest achieved 89.88% accuracy
for force location using a 0.1 cm threshold and 87.31% accuracy for force magnitude using a 0.5 N threshold, demonstrating robust performance under strict out-of-sample evaluation. To reduce dispersion and improve consistency, point-wise linearization was investigated by fitting loading (up) and unloading (down) branches; for the best-behaved
section, averaging the up/down metrics yielded a Coefficient of Determination of R2 = 0.995, a Root Mean Square Error (RMSE) of 0.49 N, and a Mean Absolute Error (MAE) of 0.44 N, supporting the feasibility of region-dependent calibration with near linear behavior in favorable areas. In addition, Kriging interpolation was employed to reconstruct 2D deformation fields from discrete measurements, enabling visualization of load propagation and contact “footprints” across the surface and providing a pathway toward continuous localization. Overall, the dissertation provides quantitative evidence
that accurate force localization and magnitude estimation can be achieved using a reduced number of strategically placed FBG sensors. The results show that the multichannel optical response contains sufficient spatial and force-related information to support reliable inference when combined with appropriate data-driven models. By integrating clustering,
supervised learning, point-wise linearization, and spatial interpolation, the proposed approach improves interpretability, accounts for region-dependent mechanical behavior, and establishes a scalable methodological basis for the progressive construction of a comprehensive pHRI testbench.
