A Soft Magnetic Tactile Sensor with Simulation-Guided History Compensation for Sub-Millimeter Contact Trajectory Tracking

Soft magnetic tactile trajectory reconstruction depends on motion and deformation history. We investigated whether an empirical dynamic simulator could inform useful refinements of an existing real-trained reconstruction chain. Virtual exploration and real-record observations motivated a function-preserving multiscale causal-history expansion trained with real data only. In two source sessions, the selected joint-objective model reduced mean command-referenced planar error from 0.436 to 0.408 mm and from 0.452 to 0.424 mm, with approximately 45% lower residual step root-mean-square (RMS). Eighteen matched runs from one initialization showed that additional data and optimization alone did not reproduce the large stability gain; a representative residual gated-recurrent-unit (GRU) model did not reproduce it either. Eight sequences from a partial repeat execution of the test plan supported joint error and stability improvement. Measurements acquired after the original model freeze supplied a separate hemispherical-indenter check: in six of 12 sequences that met a fixed baseline-comparability criterion, the common-output planar error decreased from 0.474 to 0.453 mm and the residual step RMS from 0.259 to 0.169 mm. Eight-direction holds showed lower within-hold fluctuation but not consistently lower between-direction mean bias; the median paired command-response-proxy difference was zero. Command-derived references, limited assemblies and unmeasured physical uncertainty bound stronger accuracy and latency interpretations.