Barometer-Based Tactile Sensing: Characterization, Processing, and Applications for Dynamic Manipulation

Abstract

Reliable tactile feedback is essential for robotic systems to interact effectively with their environments, especially in dynamic manipulation tasks where detecting contact onset, direction, and force is critical for control and planning. This thesis advances the development of barometer-based tactile sensors for low-force interactions, building upon prior work from the Biomimetic Robotics Lab. Previous work demonstrated that neural networks could infer contact location and three-axis contact force from barometers embedded within an elastomer. However, these models did not account for the viscoelastic behavior of the elastomer, which degrades sensor repeatability and bandwidth. To address these limitations, this thesis introduces a recurrent neural network (RNN) architecture that captures viscoelastic transients in the sensor response. The proposed methods are evaluated on two sensor geometries: a spherical sensor and a slimmer ellipsoid variant. An automated data collection pipeline is developed to generate temporally-continuous, uniformly sampled datasets across the sensor surface. RNN models trained on this data show that temporal modeling improves force prediction accuracy across both designs. To improve angle prediction accuracy, a binning strategy is used to enforce a uniform prior over contact orientations. The resulting "Binned RNN" neural networks are small-scale and demonstrate high sensitivity, enabling responsive tactile feedback. The utility of these tactile sensors is demonstrated by integrating the sensors onto a dexterous two-finger gripper and performing light grasping and estimation of object reorientation using solely tactile measurements. This work shows that accounting for viscoelastic effects through informed sampling and temporal modeling enhances the practical performance of elastomer-based tactile sensors in robotic systems.