Improving Data-Driven Contact Localization and Force Estimation for Barometric Tactile Sensors

Abstract

Barometric tactile sensors offer a cheap, robust, and customizable means for robots to perceive the world. Central to their operation are models that extract useful information from the sensors’ raw pressure readings. In this work, I focus on improving data-driven methods for single-point contact localization and force estimation using a previously presented three-quarter sphere barometric tactile sensor. To allow modeling of time-dependent effects in the sensor material, I introduce a multi-threaded data collection system that captures ground truth contact and sensor data at exactly 100 Hz. I construct both feed-forward and recurrent networks using this data, finding that a recurrent network achieves a 15% lower mean absolute error for angular contact localization on the sphere compared to prior methods. The recurrent architecture’s computational efficiency ensures that the architecture can still run within the constraints of the sensors’ microcontroller. Despite this improvement, I find that more expressive models such as LSTMs tend to overfit on the collected data and physical phenomena observed during deployment were not well represented by the training metrics. To better understand the extent that these data-driven methods alone can improve sensor performance, I shift focus away from the modeling and analyze the physical sensor instead. I find that viscous effects in the sensor can render the prediction task unlearnable without historical data and that thermal effects introduce a train-test distribution shift. Finally, I discuss design criteria for a theoretical future barometric tactile sensor that may mitigate the effects found during my modeling and analysis.