Improving Tactile Sensing for Robotic Grasping

Abstract

Enabling robots to execute reliable grasping and manipulation tasks will significantly expand the activities robots can perform. Some information that is crucial for guaranteeing grasp stability can only be obtained through tactile sensing and inaccessible by other sensing modalities. Examples include contact forces, vibration, slip onset, texture, and object mass distribution. While an increasing number of commercial and research robotic hands have embedded tactile sensors, the understanding of the behavior and limitation of tactile sensing has been limited and theoretical. In this thesis, we conducted one of the first large trial grasping experiments. We observed that tactile sensor signals may internalize a combination of uncertainties and errors in the task and throughout the robotic system, therefore exhibiting variations in the signals that are not easily separable and surmountable. To establish the fundamental information needed from the tactile sensors, we used analytical grasp stability models to identify physically intuitive and universal parameters that contact sensors should provide regardless of the underlying sensing mechanism. The same analyses can also calculate a given sensor suite's reliability and limitation when used for grasping and manipulation tasks. If used for design, our analysis can help establish quantitative criteria for both tactile sensor and system integration. Recognizing there are limitations in both pure-model and pure-data-driven learning approaches for any system dealing with real-world variations, we proposed a hybrid method of combining data-fitted, physically-inspired parametric models with nonparametric residual machine learning. We tested our hybrid approach by calculating the previous-identified physically-intuitive parameters from raw sensor data, and showed that the hybrid method out-performed both pure-model and pure-data-driven learning methods in terms of data-efficiency and generalizability.