Robotic Reproduction of Therapeutic Massage via Adaptive Admittance Control and Single-Expert Kinematic-Dynamic Modeling.

Abstract

ObjectiveThis study tested whether therapeutic massage techniques demonstrated by a single experienced practitioner can be encoded as structured kinematic-dynamic signatures and reproduced robotically under adaptive admittance control.MethodsSingle-expert manual demonstrations of four mechanically distinct techniques - beating, pressing, pushing, and vibrating - were synchronously recorded using optical motion capture and tri-axial force sensing. Each technique was acquired as a 30-s continuous trial; after predefined screening, one stable analysis segment was retained per technique (Beat 8 s, Press 20 s, Push 20 s, and Vibrate 5 s). The retained demonstrations were characterized by active Cartesian degrees of freedom, movement excursion and timing, loading direction, force amplitude, and dominant frequency. These descriptors were mapped to technique-specific nominal robot trajectories, while adaptive admittance generated contact-direction corrections during execution. Reproduction fidelity was summarized descriptively by DTW-aligned waveform correlation, mean-force bias, relative mean-force bias, peak-force error, and dominant-frequency difference.ResultsAfter DTW alignment, robotic executions preserved force waveform morphology with high descriptive similarity across techniques (Pearson r = 0.858-0.925; Spearman rho = 0.817-0.907). Descriptive error metrics identified two technique-dependent mismatch patterns. Press and Push retained the principal cyclic or loading morphology but showed marked mean-force bias (+8.25 N, +33.55% for Press; -23.38 N, -43.66% for Push). Beat and Vibrate showed smaller absolute mean-force offsets (+1.79 N and -1.54 N, respectively), although Beat still exhibited a relative mean-force bias of +21.39%. Beat and Vibrate also showed larger dominant-frequency differences (-1.25 and -1.80 Hz, respectively). Peak-force error remained within ±2.11% for all four techniques.ConclusionUnder controlled biomimetic-interface conditions, single-expert massage demonstrations could be encoded as interpretable kinematic-dynamic signatures and reproduced robotically with high temporal waveform similarity, although force-amplitude and rhythm reproduction remained technique dependent. These findings support the feasibility of the expert-to-robot representation and execution framework. Broader reproducibility and clinical relevance require validation with multiple practitioners, repeated trials, dedicated controller-performance experiments, and human participants.