NRI: FND: Robust Grasping by Integrating Machine Learning with Physical Models

点击翻译按钮获取中文摘要

This project aims to make robot grasping reliable, so that robots can pick up a wide variety of objects with minimal risk of dropping them. Except for carefully-controlled settings like factories, even the best current robotic manipulation systems drop over 5% of the objects they attempt to grasp and lift. Reliable grasping would open a vast range of practical applications in many areas of society. For example, household assistance for the elderly and disabled requires the ability to grasp and move objects, but robots that drop many objects would not be useful for many users. In addition to personal assistance functions, commercial applications that require reliable grasping range from warehouses and flexible manufacturing to food service and cleaning. Contact sensing is essential for reliable robotic grasping in unstructured environments, but existing methods for using sensor signals have not been effective. The two prior approaches to creating reliable grasping using contact sensing were physics-based grasp analysis and black-box machine learning algorithms, and neither has proved sufficient in terms of accuracy and reliability. It is not clear whether the limitations were due to deficiencies in the physical models, learning algorithms, or sensors. This project will provide the first characterization of contact sensors in carefully-controlled grasping experiments, with independent sensors providing a gold standard for comparison. The project will use the data from those experiments to investigate merging the physical modeling and machine learning approaches. By providing an improved understanding of their properties in predicting grasp stability it will allow the joining of both approaches in a way that maintains the best properties of each: interpretability and generalizability from physical modeling and flexibility and accuracy from machine learning. By pushing prediction accuracy and reliability to its peak, the project will also characterize the limits on sensor performance and define, for the first time, the needed sensor information in terms of task requirements. The results will serve as the basis for grasping and manipulation systems for unstructured environments and spell out the tradeoffs among sensor hardware, model-based signal interpretation and control, and data-driven signal interpretation and control. The successful integration of model-based and data-driven approaches may serve as an exemplar for many other fields where models are useful but incomplete.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.