Fig. 2: Marker-to-marker translation and spatiotemporal force prediction.
From: Training tactile sensors to learn force sensing from each other
A Homogeneous translation demonstrated with GelSight sensors featuring varying marker patterns. B Heterogeneous translation demonstrated with GelSight, TacTip, and uSkin sensors. C Marker-to-marker translation (M2M) model. The model takes deformed images from calibrated sensors as input, conditioned on reference images from uncalibrated sensors, to synthesize deformed images that mimic the response of the uncalibrated sensors. D Spatiotemporal force prediction model. The network processes sequential contact images to predict three-axis forces, leveraging a spatiotemporal module to enhance accuracy.
